A radar emitter individual identification method based on deep metric learning

Abstract

The application discloses a radar emitter individual identification method based on deep metric learning, and relates to the field of radar signal processing.The method is as follows: firstly, known original radar pulse waveform data is processed, single pulse data is read in sequence, initial intra-pulse features such as pulse envelope, instantaneous frequency and spectrum are extracted, then normalization, truncation and multi-feature fusion are further processed to generate training samples; then a deep neural network architecture is constructed, and a metric learning method is used to optimize the network feature extraction of individuals, so that the intra-class distance of radar individual features is reduced, and the inter-class distance is increased; finally, the network model is used to extract individual features of test data in sequence, and the identification result is obtained by inputting the classifier.Compared with the traditional radar emitter individual identification method, the method has the characteristics of fast processing speed, high feature extraction accuracy and good identification effect, and can effectively improve the identification performance of unknown radar individuals.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of radar signal processing, and particularly relates to a radar emitter individual identification method based on deep metric learning. BACKGROUND

[0002] Radar emitter individual identification is one of main functions of battlefield electromagnetic situation awareness and electronic reconnaissance equipment, and the technology is a technology for distinguishing radar emitter individuals by using unintentional modulation characteristics of radar emission signals. Through individual identification, individual distinction of radar emitters and identification of platforms thereof can be realized, and intelligence support can be provided for electronic countermeasures. After long-time development, the radar emitter individual identification technology has been widely applied to combat scenes such as key target monitoring and terminal network verification in special scenes.

[0003] However, the current radar emitter individual identification problem is faced with the following problems:

[0004] 1. The individual identification mode through traditional radar signal parameter extraction and comparison has low precision and slow speed, and is difficult to adapt to the radar emitter analysis and identification requirements in the modern dense and complex electromagnetic environment.

[0005] 2. With continuous updating of radar systems, electronic reconnaissance needs to face more open and non-static scenes, and the radar individual identification technology also needs to be optimized and updated to adapt to new scenes.

[0006] Under this condition, a new radar emitter individual identification method is needed, which faces new scenes and can improve feature precision and identification accuracy. SUMMARY

[0007] The application aims to provide a radar emitter individual identification method which has fast processing speed, high feature extraction precision, good identification effect and can cope with open scenes.

[0008] The technical solution for realizing the application is that the application provides a radar emitter individual identification method based on deep metric learning, and the method comprises the following steps:

[0009] Step 1: Preprocessing known original radar pulse waveform data, sequentially reading waveform data of each pulse, and determining intrapulse features such as pulse envelope, instantaneous frequency and spectrum;

[0010] Step 2: Normalizing, intercepting and multi-feature fusion processing based on the envelope front edge, instantaneous frequency and spectrum features of each single pulse to generate initial training sample data;

[0011] Step 3: Constructing a neural network model, determining the way of data input to the network, and determining the network architecture.

[0012] Step 4, design the loss function of the network based on the deep metric learning method, optimize the network feature distribution of the radar individual through metric learning, and train a converged network model;

[0013] Step 5, determine the single pulse fusion features of unknown radar pulse waveform data, and use the trained network model to extract individual feature vectors. The extracted vectors pass through the classification layer and output the individual identification result.

[0014] Optionally, the known original radar pulse waveform data in step 1 is preprocessed, and the waveform data of each pulse is read in sequence to determine the intra-pulse features including pulse envelope, instantaneous frequency, and spectrum.

[0015] Step 1.1, extract the envelope sequence of the original pulse waveform data:

[0016] Suppose the original waveform sequence of each pulse is r(n), 1≤n≤N, where N is the number of sampling points in a single pulse; determine the Hilbert transform of the waveform sequence Let the analytic signal sequence be The amplitude of z(n) is That is, the envelope of the original pulse waveform sequence;

[0017] Step 1.2, extract the instantaneous frequency of the original pulse waveform data:

[0018] On the basis of obtaining the analytic signal, determine the phase sequence of the signal: Where N is the number of sampling points in a single pulse; let the phase sequence be derived with respect to discrete time n to obtain the instantaneous frequency sequence

[0019] Step 1.3, extract the spectrum of the original pulse waveform data:

[0020] Discrete Fourier transform DFT is performed on the original waveform sequence r(n):

[0021]

[0022] Where N is the number of sampling points in a single pulse.

[0023] Optionally, step 2 normalizes, truncates, and multi-feature fusion processes based on the envelope, instantaneous frequency, and spectrum features of each single pulse to generate initial training sample data, including:

[0024] Step 2.1, for the envelope sequence a(n), the instantaneous frequency sequence and the spectrum sequence X(k) of each radar pulse, perform peak normalization processing to obtain the normalized corresponding feature sequence

[0025] Step 2.2, normalizing each feature sequence of each pulse Each 100-point feature fusion is generated, and the initial feature with a dimension of 300 is generated;

[0026] Step 2.3, the generated initial feature is saved as byte data according to the format of 16 bits per sampling point, as a training sample.

[0027] Optionally, step 3 constructs a neural network model, determines the way data is input to the network, and determines the network architecture, including:

[0028] A one-dimensional convolutional neural network is selected to build the network architecture: the number of input nodes of the network is 300, corresponding to the dimension of the fused features; the network part contains 5 convolutional blocks, each containing 2 convolutional layers and a pooling layer, and finally a fully connected layer is used for feature output;

[0029] The fully connected layer contains 4 layers, which are 128-32-2-Num; the 2-node output of the prelogits layer, which is the second-to-last layer, is used for dimension reduction and visualization of the extracted network features, and Num represents the number of known radar individual categories.

[0030] Optionally, step 4 designs a loss function for the network based on the deep metric learning method, optimizes the network feature distribution of the radar individual through metric learning, and trains a converged network model, including:

[0031] Step 4.1, the loss function is determined as the combination of the cross-entropy loss function of multi-classification and the center loss function in metric learning; the combination coefficient is λ, and the final loss L = L S + λL C , where L S is the cross-entropy loss function, L C is the center loss function, and λ is a coefficient that can control the aggregation degree of the same data; the value range of λ is set to [0, 0.1] with a step size of 0.01.

[0032] [0, 0.1], with a step size of 0.01.

[0033] Step 4.2, using L as the loss function, combining the training sample and the network structure, setting the number of training rounds to Num*100, where Num represents the number of known radar individuals; iterating λ in the range of [0, 0.1] with a step size of 0.01 until a converged classification model is trained and the model is saved.

[0034] Step 4.3, in the models corresponding to different λ values, select the model with the maximum inter-class distance and the minimum intra-class distance as the final individual identification model.

[0035] Compared with the prior art, the present application has the following advantages: (1) the constructed individual identification model of the deep neural network is established on the basis of a single pulse, taking intra-pulse features as initial input, so that the processing speed is faster and the real-time performance of individual identification is improved; (2) the main intra-pulse features such as envelope front, instantaneous frequency and spectrum are extracted from the original radar pulse waveform data for multi-feature fusion, so that the wide applicability and complementarity of individual network features are improved, which is beneficial to improving the individual identification accuracy; (3) the metric learning method is used to optimize the individual feature distribution in the embedding space, so that the embedding space feature distance of the same radar individual is reduced, and the embedding space feature distance of different radar individuals is expanded, which not only improves the feature extraction accuracy and recognition accuracy, but also improves the recognition effect in an open scene. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is a processing flow chart of a radar emitter individual identification method based on deep metric learning.

[0037] Figure 2 It is a network structure schematic diagram.

[0038] Figure 3 It is a comparison chart of training effects of a deep metric learning method and a conventional deep learning method.

[0039] Figure 4 It is a comparison chart of individual identification maximum classification probabilities of two network models. DETAILED DESCRIPTION

[0040] With the rapid development of deep learning technology, many researches have begun to apply deep learning to the field of radar emitter individual identification. Deep learning mainly uses convolutional neural networks as the architecture, and metric learning is to study how to learn a distance function on a specific task, so that the distance function can help the nearest neighbor-based algorithm to achieve better performance. The present application combines deep learning and metric learning methods, and proposes a radar emitter individual identification method based on deep metric learning.

[0041] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0042] In combination with Figure 1 , the present application is a radar emitter individual identification method based on deep metric learning, and the specific implementation steps are as follows:

[0043] Step 1, pre-process the known original radar pulse waveform data, read the waveform data of each pulse in turn, calculate the intra-pulse features such as pulse envelope, instantaneous frequency and spectrum, and establish the entire identification process on the basis of a single pulse, which is as follows:

[0044] Step 1.1, extract the envelope sequence of the original pulse waveform data:

[0045] Assume the original waveform sequence of each pulse is r(n), 1≤n≤N, where N is the number of sampling points within a single pulse. Determine the Hilbert transform of the waveform sequence Let the analytic signal sequence be The amplitude of z(n) is That is, the envelope of the original pulse waveform sequence.

[0046] Step 1.2, extract the instantaneous frequency of the original pulse waveform data:

[0047] On the basis of obtaining the analytic signal, determine the phase sequence of the signal: Where N is the number of sampling points within a single pulse. Let the phase sequence be derived with respect to discrete time n, and the instantaneous frequency sequence

[0048] Step 1.3, extract the frequency spectrum of the original pulse waveform data:

[0049] Discrete Fourier transform (DFT) is performed on the original waveform sequence r(n):

[0050]

[0051] Where N is the number of sampling points within a single pulse.

[0052] Step 2, based on the envelope front, instantaneous frequency and spectrum characteristics of each single pulse, perform normalization, truncation and multi-feature fusion processing to generate initial training sample data, so that the initial input features of the network have wide applicability and complementarity, as follows:

[0053] Step 2.1, for the envelope sequence a(n), instantaneous frequency sequence and spectrum sequence X(k) of each radar pulse, perform peak normalization processing to obtain the normalized corresponding feature sequence

[0054] Step 2.2, for each normalized feature sequence of each pulse Each is truncated to 100 points for feature fusion, generating an initial feature with a dimension of 300.

[0055] Step 2.3, save the generated new initial feature as byte data according to the format of 16 bits per sampling point, as a training sample.

[0056] Step 3, construct a neural network model, determine the way data is input to the network, and design the network architecture, as follows:

[0057] A one-dimensional convolutional neural network is selected to build the network architecture: the number of input nodes of the network is 300, which corresponds to the dimension of the fused features. The network part contains 5 convolutional blocks, each of which contains 2 convolutional layers and a pooling layer, and finally a fully connected layer is used for feature output. The fully connected layer contains 4 layers, which are 128-32-2-Num. The 2-node output of the second-to-last layer prelogits is used to reduce the dimension of the extracted network features for visualization, and Num represents the number of known radar individual categories.

[0058] Step 4, design the loss function of the network based on the deep metric learning method, optimize the network feature distribution of the radar individual through metric learning, train a converged network model, so that the feature distance of the same radar individual is reduced, and the feature distance of different radar individuals is expanded, as follows:

[0059] Step 4.1, the loss function is determined as the combination of the cross-entropy loss function of multi-classification and the center loss function in metric learning, and the combination coefficient is λ, and the final loss L=L S +λL C , wherein L S is the cross-entropy loss function, L C is the center loss function, and λ is a coefficient that can control the aggregation degree of the same data. The value range of λ is set to [0, 0.1], with a change step of 0.01.

[0060] [0, 0.1], with a change step of 0.01.

[0061] Step 4.2, taking L as the loss function, combining the training samples and the network structure, setting the training round number to Num*100, wherein Num represents the number of known radar individual categories. Iterating λ in the range of [0, 0.1] with a step of 0.01 until a converged classification model is trained and saved, so that the feature vector distance of the same radar individual is reduced and the feature vector distance of different radar individuals is expanded on the output feature of the prelogits layer of the network.

[0062] Step 4.3, in the models corresponding to different λ values, select the model with the maximum inter-class distance and the minimum intra-class distance as the final individual identification model.

[0063] Step 5, calculate the single-pulse fusion features of unknown radar pulse waveform data, and use the trained network model to extract individual feature vectors. The extracted vectors are passed through the classification layer to output the individual identification result.

[0064] The application will be further described in detail in combination with the drawings and specific embodiments.

[0065] Embodiment 1

[0066] 1. Experimental conditions:

[0067] In this experiment, 20 groups of pulse waveform data of 8 single-stage oscillating radar radiation source individuals are selected as training data. The pulse waveform data is binary data, and the sampling point data of each pulse is stored in sequence. Each group of pulse waveform data contains 2000 radar pulses. The 20 groups of waveform data of the training data contain 8 navigation radar individuals. Each group of data corresponds to a radar individual. At the same time, 40 groups of pulse waveform data of radar radiation sources are selected as test data, each group of data corresponds to a radar individual, and in addition to the 8 radar individuals in the training data, the test data also contains 13 groups of waveform data of unknown individuals.

[0068] Hardware platform: server, CPU is Intel Xeon(R) E5-2640 v3, main frequency is 2.6 GHz, GPU is NVIDIA GeForce RTX 3060;

[0069] Operating system: Ubuntu 18.04.6 LTS;

[0070] Development tools: Python 3.7, torch 1.12.0, torchvision 0.13.0.

[0071] 2. Experimental content:

[0072] Under the above experimental conditions, the present application is used to first preprocess all pulse waveform data, extract initial features such as pulse front and instantaneous frequency corresponding to each pulse in the data, and label the radar radiation source individual to generate data samples, construct a deep neural network to train a radar radiation source individual identification model, use a metric learning method to optimize the network model, optimize the individual feature space, and train a converged deep metric learning network model. The training effect of the data using the network without using the metric learning method and the network using the metric learning method is compared as shown in Figure 3 The left side is the extraction effect of the reduced dimension feature of the prelogits layer of the deep network trained for 300 rounds without using the metric learning method, and the right side is the extraction effect of the reduced dimension feature of the prelogits layer of the deep network trained for 300 rounds using the metric learning method. For the test samples of 13 unknown individuals, the maximum classification probability output by the traditional deep network model and the deep network model using the metric learning method is as shown in Figure 4 .

[0073] 3. Result analysis:

[0074] From Figure 3It can be seen that the feature distribution of the radar individual feature space can be obviously optimized after using the metric learning method, and the distinguishing effect of the individual features extracted by the network is improved. From the specific individual recognition result, the deep metric network model also achieves better recognition accuracy. Figure 4 The horizontal axis is the test sample number of the unknown individual, the vertical axis is the maximum classification probability of the classification into 8 known radar individuals, the broken line of the star point represents the maximum classification probability output by the traditional deep network, and the broken line of the origin represents the maximum classification probability output by the deep metric learning network in the application. From the figure, Figure 4 It can be seen that the application can significantly reduce the probability of each unknown individual being classified into a known individual, and improve the individual recognition effect in an open scene.

[0075] Compared with the prior art, the application has the following advantages: (1) the deep neural network individual recognition model is established on the basis of a single pulse, and the intra-pulse features are used as the initial input, so that the processing speed is faster, and the real-time performance of individual recognition is improved; (2) the envelope front, instantaneous frequency, spectrum and other main intra-pulse features are extracted from the original radar pulse waveform data for multi-feature fusion, so that the wide applicability and complementarity of the individual network features are improved, and the individual recognition accuracy is improved; (3) the metric learning method is used to optimize the individual feature distribution of the embedding space, so that the embedding space feature distance of the same radar individual is reduced, and the embedding space feature distance of different radar individuals is expanded, so that the feature extraction accuracy and the recognition accuracy are improved, and the recognition effect in an open scene is improved.

Claims

1. A radar emitter individual identification method based on deep metric learning, characterized in that: The method comprises: Step 1: Preprocess the known original radar pulse waveform data, read the waveform data of each pulse in turn, and determine the pulse characteristics including pulse envelope, instantaneous frequency, and spectrum; Step 2: Normalize, truncate, and fuse multiple features based on the envelope leading edge, instantaneous frequency, and spectrum characteristics of each single pulse to generate initial training sample data; Step 3: Build a neural network model, determine how to input data into the network, and determine the network architecture; Step 4: Design the network loss function based on the deep metric learning method, optimize the network feature distribution of individual radars through metric learning, and train a converged network model; Step 5: Determine the single pulse fusion features of the unknown radar pulse waveform data, and use the trained network model to extract the individual feature vectors. Pass the extracted vectors through the classification layer and output the individual recognition results. In step 2, the envelope, instantaneous frequency, and spectrum characteristics of each single pulse are normalized, intercepted, and multi-feature fusion processed to generate initial training sample data, including: Step 2.1: For each radar pulse envelope sequence a(n), instantaneous frequency sequence The spectrum sequence X(k) is subjected to peak normalization to obtain the corresponding feature sequence after normalization Step 2.2: Normalize the characteristic sequences of each pulse 100 points are intercepted for feature fusion to generate initial features with a dimension of 300; Step 2.3: Save the generated initial features as byte data in the format of 16 bits per sampling point as training samples; The loss function of the network is designed based on the deep metric learning method described in step 4. The network feature distribution of individual radars is optimized through metric learning to train a converged network model, including: Step 4.1: Determine the loss function as a combination of the cross entropy loss function for multi-classification and the center loss function in metric learning; the combination coefficient is λ, and the final loss L = L S +λL C , where L S That is the cross entropy loss function, L C is the center loss function, λ is the coefficient that controls the degree of aggregation of similar data; the value interval of λ is set to [0, 0.1], and the change step is 0.01; Step 4.2: Use L as the loss function, combine the training samples and network structure, set the number of training rounds to Num*100, where Num represents the type of known radar individuals; iterate λ in the range of [0, 0.1] with a step size of 0.01 until a converged classification model is trained and saved; Step 4.3: Among the models corresponding to different λ values, the model with the largest inter-class distance and the smallest intra-class distance is selected as the final individual recognition model.

2. The radar emitter individual identification method based on deep metric learning according to claim 1 is characterized in that: In step 1, the known raw radar pulse waveform data is preprocessed, and the waveform data of each pulse is read in sequence to determine the pulse envelope, instantaneous frequency, and spectrum. The following are the pulse characteristics: Step 1.1: Extract the envelope sequence of the original pulse waveform data: Assume that the original waveform sequence of each pulse is r(n), 1≤n≤N, where N is the number of sampling points in a single pulse; determine the Hilbert transform of the waveform sequence Let the parsed signal sequence Then the amplitude of z(n) That is, the envelope of the original pulse waveform sequence; Step 1.2: Extract the instantaneous frequency of the original pulse waveform data: Based on the obtained analytical signal, the phase sequence of the signal is determined: Where N is the number of sampling points in a single pulse; let the phase sequence be differentiated with respect to discrete time n to obtain the instantaneous frequency sequence Step 1.3: Extract the spectrum of the original pulse waveform data: Perform discrete Fourier transform DFT on the original waveform sequence r(n): in N is the number of sampling points in a single pulse.

3. The radar emitter individual identification method based on deep metric learning according to claim 1 is characterized in that: Step 3 builds a neural network model, determines how data is input into the network, and determines the network architecture, including: A one-dimensional convolutional neural network was selected to build the network architecture. The number of input nodes in the network was 300, corresponding to the dimension of the fused features. The network consisted of five convolutional blocks, each of which contained two convolutional layers and one pooling layer, and finally a fully connected layer for feature output. The fully connected layer consists of 4 layers, which are 128-32-2-Num. The 2-node output of the penultimate prelogits layer is used to visualize the dimensionality reduction of network extracted features. Num represents the type of known radar individual classification.

Images