A model embedding based emitter identification method

Abstract

The application provides a model embedding-based radiation source identification method, different modal data are jointly input, an end-to-end multi-modal identification framework is trained, parameter sharing between different modal models is realized, compared with multiple modal corresponding to multiple network identifications, the network scale is obviously reduced, the convergence speed is improved, and in the process of sharing parameters of multiple modes, the network obtains more common information between modes. Meanwhile, after the end-to-end identification framework, a fusion algorithm based on the Bayes theory is provided, compared with using a single mode as the network output, since the output after fusion comes from different modes, the confidence of the result is higher, therefore, the fused output is used as the decision of the radiation source individual identification, the iteration number required during network convergence is reduced, and the effect of optimizing the network performance is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electromagnetic signal processing, in particular to the technical field of emitter identification. BACKGROUND

[0002] In recent years, with the rapid development of electronic technology, the intelligent application of electromagnetic spectrum has shown explosive growth, providing convenience for people's life. However, in fact, the electromagnetic data obtained in a complex environment has the problems of large data volume and few effective samples of the target. The multi-modal emitter threat detection and individual identification technology is a difficulty in processing massive electromagnetic data. Since the signal characteristics of the electromagnetic target cannot be known in advance, the classifier design is complex and the recognition rate is low. In addition, for new electromagnetic targets, it is difficult to obtain enough training samples, and a classification model with high classification precision cannot be trained. Therefore, how to accurately re-identify the multi-modal emitter identification is crucial for obtaining the dominant right on the battlefield.

[0003] Emitter identification identifies the type, working mode and other target information of the emitter by measuring the radar repetition period, carrier frequency, pulse width, amplitude and other signal parameters, and processing. With the continuous development and progress of modern information technology, the traditional emitter identification technology has been unable to meet the identification requirements of various complex signals. With the deep learning technology, the feature extraction work can be automated, and more essential features of the data can be extracted. In the case of sufficient number of samples, it can have good recognition accuracy and robustness.

[0004] We usually define the data of the same target from different sensors as multi-modal data. Generally, due to the scattered distribution of communication data of the emitter in time, it is difficult to collect, and multiple collection devices are arranged in a certain area to obtain multi-source heterogeneous data of the emitter. This kind of emitter data can also be called multi-modal emitter data. For multi-modal emitter data, due to the influence of the fingerprint features of the collection device, the general network does not have good generalization performance for single modal emitter training, so multiple networks need to be trained, which leads to the disadvantages of large network size and redundant model quantity. SUMMARY

[0005] The technical problem to be solved by the present application is that the quantity of single identification parameters for multi-modal is too large, and the number of training times is too much. A radiation source individual identification method is provided, which can improve the individual accuracy of radiation source identification in a multi-modal scenario, improve the performance of neural networks, and reduce the size of the network.

[0006] The technical scheme adopted by the present application to solve the above technical problems is a radiation source identification method based on model embedding, comprising the following steps:

[0007] Training steps:

[0008] 1) Time-frequency data preprocessing is performed on electromagnetic signals from each radiation source;

[0009] 2) Determine the type of each modality of the multi-modal, and extract different modal time-frequency features from the preprocessed time-frequency data to obtain a sample feature matrix;

[0010] 3) Data labeling is performed on the sample feature matrix to obtain a sample feature matrix and corresponding radiation source class label, and data set division is performed on the data labeled sample feature matrix to obtain a training set and a test set;

[0011] 4) Embedding and connecting w neural network models corresponding to the modality type according to the number of modality types, inputting the training set into the w neural network models, inputting a training sample in the training set into one of the w neural network models, and completing pre-training through parameter sharing of the w neural network models;

[0012] 5) Input the test set into the pre-trained w neural network models for radiation source identification:

[0013] 5-1) When the test set is input into the neural network model, it is ensured that each batch of w test samples entering the neural network model belongs to the same radiation source class; the probability matrix output by each of the w neural network models, each row of the probability matrix corresponds to a test sample, and each column is the identification probability of a different radiation source class;

[0014] 5-2) Fuse each row of the probability matrix output by the w neural network models of different modalities to output the final fusion probability:

[0015]

[0016] P(i) represents the fusion probability of the ith row, when each batch of w test samples entering the neural network model belongs to the same radiation source class, the same row of the probability matrix output by the w neural network models belongs to the same radiation source class, and p i1 represents the identification probability of the ith row of the probability matrix of the first neural network model, p iw represents the identification probability of the ith row of the probability matrix of the wth neural network model.

[0017] 5-3) Determine whether the identification accuracy meets the network convergence condition according to the fusion probability output by the w neural network models, if yes, save the model parameters of the w neural network models, and the training process is ended, otherwise, adjust the hyperparameters of the w neural network models and return to step 5-1);

[0018] Identification steps:

[0019] The electromagnetic signal of the to-be-identified radiation source is subjected to time-frequency data preprocessing, time-frequency features are extracted, and then are input into the w neural network models which are trained in parallel, and the radiation source type is identified according to the fusion probability output by the w neural network models.

[0020] The application realizes parameter sharing between different modal models through an end-to-end identification framework based on multi-modal radiation source individual signal identification, shares model parameters after multi-modal time-frequency features of target signals, and performs Bayesian fusion at the decision layer on the basis of the discrimination decision after the model parameter sharing. Different modal data are jointly input, trained through the end-to-end multi-modal identification framework, realize parameter sharing between different modal models, and the network size is significantly reduced compared with multiple modal network identification, the convergence speed is improved, and in the process of sharing parameters of multiple modalities, the network obtains more common information between modalities. The commonality between different modalities is learned, and deeper radiation source features are extracted. This is also the subtle feature information of the radiation source individual that we need, and the recognition accuracy is higher than that of single recognition, so that the recognition result is higher.

[0021] Meanwhile, based on the output of the network, the application provides a fusion algorithm based on Bayesian theory after the end-to-end identification framework. Compared with using single modal as network output, since the output after fusion comes from different modalities, the confidence of the result is higher, so the output after fusion is used as the decision of radiation source individual identification, the iteration number required during network convergence is reduced, and the effect of optimizing network performance is achieved.

[0022] The application has the beneficial effects that parameter sharing between different modal models is realized, the network size is reduced, the convergence speed is accelerated, the network performance is optimized, and the accuracy of radiation source individual identification is improved through related calculation of the network output through the fusion algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A radiation source individual identification process based on model embedding in two modal scenarios is shown in the embodiment.

[0024] Figure 2 A multi-modal identification framework based on model embedding is shown in the embodiment.

[0025] Figure 3Figures of time-domain waveforms of original signals of radiation sources for the embodiment; (a) is the original time-domain waveform of radiation source 1 of receiver a, (b) is the original time-domain waveform of radiation source 2 of receiver a, (c) is the original time-domain waveform of radiation source 3 of receiver a, (d) is the original time-domain waveform of radiation source 4 of receiver a, (e) is the original time-domain waveform of radiation source 1 of receiver b, (f) is the original time-domain waveform of radiation source 2 of receiver b, (g) is the original time-domain waveform of radiation source 3 of receiver b, and (h) is the original time-domain waveform of radiation source 4 of receiver b.

[0026] Figure 4 Figures of time-frequency diagrams of radiation source 1 in different transform domains for the embodiment; (a) is the time-frequency diagram of radiation source 1 of receiver a, and (b) is the time-frequency diagram of radiation source 1 of receiver b.

[0027] Figures of time-frequency diagrams of radiation source 1 in different transform domains for the embodiment; (a) is the time-frequency diagram of radiation source 1 of receiver a, and (b) is the time-frequency diagram of radiation source 1 of receiver b. DETAILED DESCRIPTION

[0028] The multi-modal of the present application refers to two or more modalities. The embodiment takes two modalities as an example.

[0029] Figure 1 is a model embedding-based radiation source individual identification process in the two-modality scenario of the embodiment, which mainly includes the following steps:

[0030] 1. Data preprocessing;

[0031] 2. Time-frequency feature extraction of different modalities;

[0032] 3. Data labeling and dataset division process;

[0033] 4. Model embedding, model training, and model parameter sharing to obtain a trained network model;

[0034] 5. Inputting test data into the trained network model, outputting individual prediction results by the network model, and outputting the final radiation source individual identification result through a decision fusion algorithm based on Bayesian theory.

[0035] The data preprocessing process of step 1 includes filtering, energy detection, time slicing, and normalization processing of the original data, which is similar to the ordinary signal preprocessing method. Steps 2 and 3 are time-frequency feature extraction of different modalities and data labeling, which are the same as the conventional time-frequency feature extraction method and labeling method.

[0036] The main difference from the prior art is the model embedding of step 4 and the decision fusion technology of step 5. Through the processing of steps 4 and 5, the problem of excessive model parameters in the multi-modality-based radiation source individual identification task can be successfully alleviated, and the multi-modality-based radiation source individual identification accuracy can be improved.

[0037] The model embedding process of step 4 is mainly based on an end-to-end model, such as Figure 2 , the model framework is a dual-path end-to-end multi-modal recognition network, which learns feature representation and distance metric in an end-to-end manner while maintaining high discriminability. It includes two main components: a dual-path network composed of a modal one path and a modal two path for feature extraction and a bidirectional double-constrained ranking loss for feature learning.

[0038] Specifically, the dual-path learns multi-modal shared features by simultaneously modeling modal-specific information and modal-shared information using partially shared structures. The double-constrained loss ensures that the learned feature representation has sufficient discriminability to distinguish different radiation source signals from the two different modalities.

[0039] One path of the dual-path multi-modal recognition network receives input of one modality, and each modality input has a batch size of N time-frequency feature images. So the single input for training has 2*N time-frequency feature images. The 2*N time-frequency feature images are first input to the DenseNet, and then the output of the DenseNet is output to the batch normalization BN through the first fully connected network FC1, which is similar to the ordinary signal preprocessing method. DenseNet is a shallow feature extractor, and its weights are different to extract modal-specific information, while the weights of the embedding FC layer (feature embedding) are shared for multi-modal shared feature learning. After L2 normalization, a bidirectional double-constrained loss is introduced for network training. At the same time, the identification loss Identify Loss is further combined with the Ranking loss to improve performance.

[0040] For a given batch, it contains N modal one and N modal two data. For a modal one sample x i , its label is y i , we hope that the feature distance D between its same class sample z j is smaller than the feature distance D between its different class sample z k and a pre-defined boundary p1:

[0041]

[0042] Where the vector x and the vector z are the converged feature vectors after L2 normalization, and the Euclidean distance is used as the similarity measure. Therefore, a modal one to modal two ranking loss is given:

[0043]

[0044] Different modalities of the same sample can be considered as the same label, so we get the identityLoss as:

[0045]

[0046] The fusion algorithm of step 5 is based on Bayesian theory. We fuse the results of the two modalities output by the model. Let the two probabilities output by the model for a single target be P(a) = {p 1a ,p 2a ,…} and P(b) = {p 1b ,p 2b ,…}, where p ia represents the probability that the model obtains the target as the i-th class under the a modality, and p ib represents the probability that the model obtains the target as the i-th class under the b modality. According to the Bayesian theory formula:

[0047]

[0048] The fused probability formula is:

[0049]

[0050] where P(i) represents the probability that the target is the i-th class based on the fusion algorithm.

[0051] Next, we take a four-class multi-modal radiation source individual identification task as an example. Four existing radiation source individuals of the same type receive data from two different receivers, and their original signal time-domain waveforms are shown in Figure 3 .

[0052] Take a pulse data generated by a radiation source individual as a sample. Take 800 samples of each modality of each radiation source individual for training and 200 samples for testing. Take the features extracted by the short-time Fourier transform time-frequency method as an example:

[0053] First, perform time-frequency transformation on each sample in different modalities. Take radiation source 1 as an example. The time-frequency graphs after time-frequency analysis of different modalities are shown in Figure 4 .

[0054] The different modalities of each sample are required to obtain the same dimension complex matrix after time-frequency transformation processing (224*224 as an example). A sample obtains a four-dimensional feature with a dimension of [1, 2, 224, 224, 2] after transformation, wherein the first dimension represents the number of samples, the second dimension represents the number of modalities, the third and fourth dimensions are the length and width of the time-frequency graph, and the fifth dimension represents the number of channels. The training set of all sample data (4 radiation sources) is 4*800 = 1600, and the feature dimension is [1600, 2, 224, 224, 2]. Then, the feature is input into an end-to-end embedding model as a radiation source individual fingerprint feature for training.

[0055] The neural network model adopts Densenet121, and the objective function selects a combination of cross-entropy loss function, ranking loss function and identify loss function.

[0056]

[0057] wherein λ1, λ2 and λ3 are weights of the cross-entropy loss function the ranking loss function and the identify loss function.

[0058] The cross-entropy loss formula is:

[0059]

[0060] wherein y represents a true label, and represents an estimated value calculated by the network.

[0061] An Adam optimizer is used to find the optimal solution. In testing and actual application, a trained network model is loaded, test set data is input into the network, the neural network is used to classify and judge, and a classification result is output after calculating the probabilities of each class by a fusion algorithm.

[0062] The overall steps are as follows:

[0063] 1) Data preprocessing is performed on the collected original signal intermediate frequency data of the radiation source, and the specific steps are as follows:

[0064] a) Filtering is performed on the time domain intermediate frequency data of the original signal, so as to suppress the influence of noise;

[0065] b) Energy detection is performed on the filtered signal to extract valid signal intermediate frequency data;

[0066] c) IQ splitting is performed on the extracted valid data to obtain complex signal data;

[0067] ​d) The modulus of the obtained complex signal data is normalized to convert the modulus domain of the original data to

[0068] [-1,1] range, |x| max is the maximum value of the modulus of the complex signal, x * is the normalized sample data, and its normalization formula is

[0069]

[0070] e) The normalized data is sliced according to the same time domain length to form data samples for time-frequency analysis.

[0071]

[0072] 2) The same time-frequency processing operation is performed on different modal data to obtain a five-dimensional time-frequency matrix, wherein,

[0073] The first dimension represents the number of samples, the second dimension represents the number of modes, the third and fourth dimensions are the length and width of the time-frequency graph, and the fifth dimension represents the number of channels. The number of samples, length, width, and number of channels of each mode need to be the same.

[0074] 3) The sample feature matrix obtained under different time-frequency analysis methods is made into a data set under the time-frequency analysis method, and the data set is made by the same method, and the specific steps are as follows:

[0075] a) One-hot encoding is used to generate a one-dimensional label matrix corresponding to each radiation source;

[0076] b) The one-dimensional label matrix is corresponding to the sample feature matrix extracted from the time-frequency analysis;

[0077] c) The training set and the test set are divided according to 8:2;

[0078] d) The data of the training set is combined and shuffled between two modes, and the test set is combined and shuffled in the same way during the training process.

[0079] 4) Construct a neural network, and the neural network selects Densenet121, and the objective function selects the combination of cross-entropy loss function, Ranking loss function and Identify loss function, and its formula is:

[0080]

[0081] 5) The data sets obtained under different modal time-frequency analysis methods are concatenated and input into the neural network of the end-to-end multi-modal recognition framework for pre-training, and finally the training set accuracy rate a i of different modal data is recorded when the network converges. ​

[0082] 6) Test according to the converged network, input the test set into the end-to-end network for identification, and make decision fusion of the obtained different modal probabilities through the fusion algorithm, the specific steps are as follows:

[0083] a) Fix the sample reading order when reading data during the test process, so that each pair of samples entering the model is of the same category;

[0084] b) Suppose that the output of the trained model is two B*N matrices, where B is the Batch size size and N is the number of categories, then two B*N probability matrices can be obtained, where each row is a sample and each column is the probability of different categories;

[0085] c) Fuse the output probability of each row according to the following formula:

[0086]

[0087] d) Take the obtained probability as the probability output of the fusion algorithm, and record the accuracy;

[0088] 7) Adjust the neural network hyperparameters, save the model after the network converges;

[0089] 8) Test set verification, complete the individual identification of the radiation source.

Claims

1. A model embedding based emitter identification method, characterized in that, The method comprises the following steps: The training step: 1) Time-frequency data preprocessing is performed on electromagnetic signals from each radiation source; 2) Determine the type of each modality of the multi-modal, extract different modal time-frequency features from the preprocessed time-frequency data to obtain a sample feature matrix; 3) Data labeling is performed on the sample feature matrix to obtain a sample feature matrix and corresponding radiation source class label, and the sample feature matrix after data labeling is divided into a training set and a test set; 4) Embedding and connecting w neural network models corresponding to the modality type according to the number of modality types, inputting the training set into the w neural network models, inputting a training sample in the training set into one of the w neural network models, and completing pre-training through parameter sharing of the w neural network models; 5) Input the test set into the pre-trained w neural network models for radiation source identification: 5-1) When the test set is input into the neural network model, it is ensured that each batch of w test samples entering the neural network model belongs to the same radiation source category; the probability matrix output by each of the w neural network models, each row of the probability matrix corresponds to a test sample, and each column is the identification probability of a different radiation source category; 5-2) Fuse each row of the probability matrix output by the w neural network models of different modalities to output the final fusion probability: P(i) represents the fusion probability of the ith row, when each batch of w test samples entering the neural network model belongs to the same radiation source category, the same row of the probability matrix output by the w neural network models belongs to the same radiation source category, p i1 the identification probability on the ith row of the probability matrix of the first neural network model, p iw represents the identification probability on the ith row of the probability matrix of the wth neural network model; 5-3) According to the fusion probability output by the w neural network models, determine whether the identification accuracy meets the network convergence condition, if yes, save the model parameters of the w neural network models, and the training process is ended, otherwise, adjust the hyperparameters of the w neural network models and return to step 5-1); The identification step: Time-frequency data preprocessing is performed on the electromagnetic signals of the radiation source to be identified, and the time-frequency features are extracted and input into the trained w neural network models in parallel, and the radiation source type is identified according to the fusion probability output by the w neural network models.

2. The method of claim 1, wherein, In step 2), the five-dimensional sample feature matrix is obtained by extracting different modal time-frequency features from the preprocessed time-frequency data, wherein the first dimension represents the number of samples, the second dimension represents the number of modalities w, the third and fourth dimensions are the length and width of the time-frequency graph, and the fifth dimension represents the number of channels of the time-frequency data. The number of samples, the length and width of the time-frequency graph, and the number of channels of the sample feature matrix of different modalities are the same.

3. The method of claim 1, wherein, The time-frequency data preprocessing specifically includes the following steps: a) Filtering the time-domain intermediate frequency data of the electromagnetic signals from each radiation source; b) Energy detection is performed on the time-domain intermediate frequency data to extract valid signals; c) IQ demultiplexing is performed on the valid signals to obtain complex signals; d) performing a modulus normalization of the complex signal x to obtain normalized sample data x * : where |x| is the modulus of the complex signal x. max is the maximum value of the modulus of the complex signal x. e) After normalization, the sample data is sliced according to the same time domain length to form data samples for time-frequency processing as preprocessed time-frequency data.

4. The method of claim 1, wherein, Step 3) is as follows: a) One-dimensional label matrix corresponding to each radiation source is generated by using one-hot encoding; b) The one-dimensional label matrix is labeled corresponding to the sample feature matrix; c) The labeled sample feature matrix is divided into a training set and a test set according to 8:2; d) The sample feature matrix in the training set is combined and shuffled between different modalities.

5. The method of claim 1, wherein, The neural network model uses Densenet121; Objective function Using a cross-entropy loss function Ranking loss function And a combination of Identify loss function ​ wherein λ1, λ2, and λ3are the weights of the cross-entropy loss function ranking loss function and the Identify loss function loss function.

Images