Radar radiation source multi-task integrated identification method and device based on deep learning

Abstract

The invention discloses a radar radiation source multi-task integrated recognition method and device based on deep learning, and belongs to the field of radar radiation source multi-task integrated recognition, and the method comprises the steps: S1, recognition model construction: carrying out the feature representation of the model type, working mode and working system of a radar radiation source, and achieving the construction of a recognition model; s2, identification model training: constructing a training data set, training the priori knowledge model, and storing a model training hyper-parameter result, a network architecture of the identification model and a weight coefficient obtained by training in an identification knowledge file; and S3, identification model testing: loading hyper-parameters, identifying the network architecture and parameters of the model, testing the performance of the intelligent identification model by using the test data, finally identifying the data frame sequence in sequence, and storing the identification test result for subsequent evaluation. According to the method, multi-task identification of models, states and system attributes can be integrally realized, the identification accuracy is equivalent to that of an existing method, and generalization is higher.

Description

Technical Field

[0001] The present invention relates to the field of multi-task integrated recognition of radar radiation sources, and more specifically, to a method and device for multi-task integrated recognition of radar radiation sources based on deep learning. Background Art

[0002] Among the existing schemes, Xu Tao and others from the National University of Defense Technology proposed a radar emitter identification method based on a decision tree by extracting and utilizing the high-dimensional features of radar pulse intervals. This method uses the vector formed by adjacent pulse intervals as the high-dimensional features of the pulse to enhance the separability between different radar signals, and then uses a clustering method to extract this feature in the pulse train. After that, the feature is formed into a feature vector and sent to the constructed decision tree classification model. Finally, the learned model is used for the identification of radar pulse trains. This method can accurately identify the type of radar emitters with complex repetition patterns, and has a good recognition accuracy rate when there are fewer training samples. However, this method is only for the identification of radar emitter types with stable pulse repetition intervals. The identification performance of radar emitters with varying pulse repetition intervals is relatively general, and it is unable to identify information such as the working system and working mode of the radar emitter.

[0003] Pu Yunwei and others from Northeastern University have proposed an integrated deep learning recognition method that integrates the multi-domain projection features of fuzzy functions to solve the problems of poor anti-noise performance and low recognition accuracy of radar radiation source signal recognition methods in complex electromagnetic environments. This method first performs Gaussian smoothing on the fuzzy function of the signal, and selects appropriate angles from a multi-domain perspective to perform two-dimensional projection of the fuzzy function to construct a feature data set; then, multiple densely connected networks DenseNet are used as primary classifiers to train and learn three types of feature data sets respectively to obtain primary classification results; finally, the primary classification results are fused and learned through the Stacking strategy to obtain the final category information. However, this method uses the three-dimensional spatial image of the fuzzy function for feature extraction, and then uses the features for signal recognition. The feature extraction itself depends on the accuracy of the three-dimensional image construction of the fuzzy function. In practical applications, it is difficult to obtain a high-precision fuzzy function, and when the actual radiation source signal does not match the theoretical assumptions of the fuzzy function, the fuzzy function will fail, resulting in a significant reduction in the accuracy of this method. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a deep learning-based multi-task integrated identification method and device for radar radiation sources, which can realize multi-task identification of model, state, and system attributes in one integrated manner. The recognition accuracy is comparable to that of the existing methods and has stronger generalization.

[0005] The object of the present invention is achieved through the following solutions:

[0006] A multi-task integrated identification method for radar emitters based on deep learning, comprising the following steps:

[0007] S1, identification model construction: by defining the characteristic space of the radar radiation source, the type, working mode and working system of the radar radiation source are characterized to realize the identification model construction;

[0008] S2, recognition model training: construct a training data set, and then use the recognition model constructed in step S1 to train the prior knowledge model, and store the model training hyperparameter results, the network architecture of the recognition model and its trained weight coefficients in the recognition knowledge file, in preparation for the model test in step S3;

[0009] S3, recognition model test: load the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters, then use the test data to test the performance of the intelligent recognition model, and finally rely on the loaded intelligent recognition model to identify the data frame sequence in sequence, and save the recognition test results for subsequent evaluation.

[0010] Furthermore, in step S1, the feature space of the radar radiation source is defined, and the radar radiation source model type, working mode, and working system are characterized to realize the construction of the recognition model, which specifically includes the following sub-steps:

[0011] Step S11, firstly divide the signal pulse sequence of each tracking batch number in a non-overlapping manner according to a certain time length Δ Forming discrete processing frames

[0012]

[0013] Where F represents the signal pulse sequence of the radiation source tracking batch number i The number of discrete frames that can be divided; for each frame pulse sequence

[0014] Step S12, using a deep neural network to further extract the global sequence potential features x between multiple frames global_embedding The deep neural network includes convolutional neural network CNN, recurrent neural network RNN ​​and its variant networks; the specific network construction is as follows:

[0015] The CNN network uses a one-dimensional convolutional layer Conv1D to transform the dimension and expand the vector dimension, so that the network can learn more complex feature representations. The formula is as follows:

[0016] y CNN =CNN(x;θ)

[0017] Among them, x represents the input signal sequence frame, θ represents the parameters of the network, and y CNN Represents the feature vector output after processing by the CNN network;

[0018] After that, the one-dimensional convolution result is sent to the batch normalization layer to improve the training speed and stability; the variable is then sent to the maximum pooling layer to reduce the vector dimension to reduce the model calculation amount;

[0019] The above three networks constitute the identification network block, which is expressed as:

[0020] y block =P(L(A(C(x))))

[0021] Among them, C represents the convolution operation, A represents the activation function, L represents the normalization operation, P represents the pooling operation, and y block represents the feature vector output after being processed by the network block; the recognition network block will be repeated m times in the network to enhance the feature extraction capability of the network;

[0022] After that, we use RNN and its variant networks in conjunction. By sending the vector results processed by the recognition network block to two LSTM layers respectively, we make them learn the characteristics of different aspects of the sequence independently. The formula is:

[0023] h 1 =LSTM 1 (y block ; Θ 1 )

[0024] h 2 =LSTM 2 (y block ; Θ 2 )

[0025] Among them, y block is the feature vector output by the CNN network, Θ 1 and θ 2 Represent the parameters of the two LSTM networks, h 1 and h 2 are the hidden states of the two LSTM networks at the last moment of the time series; the concatenate function is used to stack the outputs of the two LSTM layers; the dual LSTM network and Concatenate operations are repeated n times to effectively represent the multi-frame context content features x global_embedding ;

[0026] Afterwards, the vector is sent to the fully connected layer for feature integration and mapping;

[0027] Finally, three different fully connected networks (FCNs) are used to extract different potential feature vectors that tend to distinguish models, states, and systems, namely:

[0028] x type =φ type (W type x global_embedding )

[0029] x mode =φ mode (W mode x global_embedding )

[0030] x class =φ class (W intent x global_embedding )

[0031] In the formula, W represents the weight matrix of the fully connected layer, and φ represents the nonlinear activation function of the output layer. Finally, the three different potential representation vectors that distinguish the model, state, and system are input into the corresponding classifiers to obtain the corresponding prediction results:

[0032]

[0033] Where W classfication is the weight matrix of the classifier.

[0034] Further, in step S2, the training data set is constructed, and the recognition model constructed in step S1 is used to train the prior knowledge model, and the model training hyperparameter results, the network architecture of the recognition model and the weight coefficients obtained by training are stored in the recognition knowledge file, in preparation for the model test in step S3, which specifically includes the following sub-steps:

[0035] S21, preprocessing the PDW sequence: first, sampling the PDW sequence to reduce the amount of data, and the sampled PDW sequence is defined as the SDW sequence; secondly, according to the timestamp in the SDW sequence data, the data frame length T and the inter-frame interval ΔT are taken to perform frame interception on the SDW sequence; according to the timestamp in the SDW sequence data, the timestamp, sequence features, and sequence labels in the same frame of SDW data are divided and stored for subsequent training;

[0036] Step S22: construct the training data set D required for training the intelligent recognition model train ; For each annotated process data file in the training file path set, according to the agreed data frame length, randomly sample training sample pairs proportional to the total length of the process data to construct the training data set D train ;

[0037]

[0038] Where N train represents the number of process data files in the training set, M i represents the number of samples collected by the i-th process data file, Δ represents the duration of the sample frame, represents the jth sample collected in the i-th file, represents the label vector of the jth sample collected in the i-th file, Indicates that the jth sample collected in the i-th file contains a pulse sequence quantity;

[0039] Step S23: training and generating the prior knowledge model required for intelligent identification of model, state and system Where θ is the model parameter, θ hp is the hyperparameter of the model; according to the multi-input multi-task learning neural network architecture built, and the multi-task optimization loss function L is set; with the training data set D train The sample pairs in are used as input to train the model parameters of the recognition model; the multi-task optimization loss function L is defined as:

[0040] L=λ type L type +λ mode L mode +λ class L class ,∑λ i =1

[0041]

[0042] In the formula, λ type , mode and λ class are the weight coefficients for model type, working mode and working system identification classification, L type , L mode and L class are the loss functions for model type, working mode and working system identification and classification, respectively, C type , C mode and C class are the number of categories of model type, working mode and working system, and are the true labels of the i-th sample in the j-th category of model type, working mode and working system, respectively. and are the predicted probabilities of the i-th sample in the j-th category of model type, working mode and working system, respectively, and n is the number of batches for training;

[0043] Step S24: first store the hyperparameters in the specified recognition knowledge file, and then store the network architecture of the aforementioned recognition model and the weight coefficients obtained by training in the specified recognition knowledge file.

[0044] Furthermore, in step S23, the model parameters of the training recognition model are specifically trained using a gradient descent training method according to a set batch size, number of iterations and optimization strategy.

[0045] Furthermore, in step S24, the hyperparameters are specifically hyperparameters required for preprocessing and data normalization of intelligent recognition.

[0046] Further, in step S3, the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters are loaded, and the performance of the intelligent recognition model is tested using the test data. Finally, the data frame sequence is sequentially recognized based on the loaded intelligent recognition model, and the recognition test results are saved for subsequent evaluation, which specifically includes the following sub-steps:

[0047] S31, model evaluation index establishment: firstly, the model evaluation index is established, and K possible model type recognition results are required to be given for each frame time f within the duration of each radar radiation source tracking batch; Indicates that the model type estimated by batch number t in frame f is k, otherwise it is 0; record a t,f Represents the radiation source associated with the tracking lot number t, and obtains the true model type indicator function When the model type of the radiation source associated with the tracking lot number t in frame f is k, its value is 1, otherwise it is 0;

[0048] The confusion matrix is ​​used to quantify the model type classification index of the electronic warfare system; the confusion matrix elements are different model types, and their values ​​are the number of true positives, true negatives, false positives and false negatives, that is,

[0049]

[0050] Among them, f t and F t are the first and last frames of the tracking activity of batch number t respectively; according to the above confusion matrix, the recognition accuracy of model type k is defined as

[0051]

[0052] Step S32, parameter model loading: load the required hyperparameters and recognition model The network architecture and its parameters are obtained; first, the intelligent recognition hyperparameters are loaded from the specified recognition knowledge file to initialize the data preprocessing and data normalization processing functions; then, the intelligent recognition model content is loaded, the neural network architecture of multi-input multi-task learning is reconstructed, and the network is initialized using the model parameters;

[0053] Step S33, specific test and result saving: relying on the loaded intelligent recognition model Identify the data frame sequence in turn, calculate the model recognition accuracy, and save the recognition test results; each discrete data frame sequence in the test set o i Each sample frame in Input into the intelligent recognition model in sequence Get the prediction results of model type, working mode and working system identification

[0054]

[0055] The predicted result and the true value label are paired The sequence is saved in the test result file corresponding to the i-th process data file for subsequent evaluation.

[0056] A multi-task integrated identification device for radar radiation sources based on deep learning comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, any of the methods described above is executed.

[0057] The beneficial effects of the present invention include:

[0058] The present invention provides a multi-task integrated recognition method based on a new representation of waveform signals. By using deep learning methods such as convolutional neural networks (DNNs), recurrent neural networks (RNNs) and their variants, the model, state and system recognition of radiation sources are realized in one piece by capturing different levels of distinguishing sequence features such as waveform style, resource scheduling and mode switching. By capturing the biased characteristics of different task features contained in the waveform sequence that can effectively distinguish model types, working states, and working systems, the recognition of different tasks of radiation sources is finally realized in one piece. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0060] Figure 1 It is a schematic diagram of the overall process of the radar radiation source multi-task integrated identification method according to an embodiment of the present invention;

[0061] Figure 2 A schematic diagram of a neural network training process according to an embodiment of the present invention;

[0062] Figure 3 A schematic diagram of a model test process according to an embodiment of the present invention;

[0063] Figure 4 This is an example diagram of a single sequence recognition result according to an embodiment of the present invention;

[0064] Figure 5 This is a statistical chart of the recognition accuracy results of an embodiment of the present invention. DETAILED DESCRIPTION

[0065] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0066] In the conception of the preferred embodiment of the present invention, the existing single attribute identification method of radar radiation source based on traditional machine learning or deep learning will greatly increase the overhead and load of system computing and storage resources when realizing multi-attribute identification, which is not conducive to the beneficial complementation between each other (such as the TAS mode is conducive to phase scanning system), and is prone to overfitting. In view of the above problems, the present invention proposes an integrated identification method of radar radiation source attributes based on deep learning in the preferred implementation scheme. The method uses a deep learning method of stacking multi-layer convolutional neural network layers, bidirectional recursive neural networks and their variant layers to realize multi-task identification of model, state, and system attributes in one. The recognition accuracy is comparable to that of existing methods, and the generalization is stronger.

[0067] The specific implementation process of the present invention is as follows:

[0068] The implementation steps of the present invention are mainly divided into three parts, namely, recognition model construction, recognition model training, and recognition model testing. First, in step S1, by defining the feature space of the radar radiation source, the radar radiation source model type, working mode, and working system are characterized to realize the recognition model construction. Secondly, in step S2, a training data set required for training the intelligent recognition model is constructed, and then the recognition model constructed in step S1 is used to train the prior knowledge model, and the model training hyperparameter results, the network architecture of the recognition model and the weight coefficients obtained by training are stored in the recognition knowledge file to prepare for the model test in step S3. Finally, in step S3, the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters are loaded, and then the performance of the intelligent recognition model is tested using the test data. Finally, the loaded intelligent recognition model is used to identify the data frame sequence in sequence, and the recognition test results are saved for subsequent evaluation. The specific implementation steps are as follows:

[0069] Step S1, identification model construction: using convolutional neural network (CNN), recurrent neural network (RNN) and their variants of deep learning methods to realize the model, state and system identification of radiation sources in one, it is necessary to first build a deep learning neural network model.

[0070] Furthermore, the step S1 includes:

[0071] Step S11, firstly divide the signal pulse sequence of each tracking batch number in a non-overlapping manner according to a certain time length Δ Forming discrete processing frames

[0072]

[0073] Where F represents the signal pulse sequence of the radiation source tracking batch number i The number of discrete frames that can be split.

[0074] For each frame pulse sequence

[0075] Step S12: using CNN network, RNN network and its variants to further extract the global sequence potential features x between multiple frames global_embedding The specific network construction is as follows:

[0076] The CNN network uses a one-dimensional convolutional layer Conv1D to transform the dimension and expand the vector dimension, so that the network can learn more complex feature representations. The formula is as follows:

[0077] y CNN =CNN(x;θ)

[0078] Among them, x represents the input signal sequence frame, θ represents the parameters of the network, and y CNN Represents the feature vector output after being processed by the CNN network.

[0079] The one-dimensional convolution result is then sent to the batch normalization layer to improve training speed and stability. The variable is then sent to the maximum pooling layer to reduce the vector dimension to reduce the amount of model calculations.

[0080] The above three networks constitute the identification network block, which can be expressed as:

[0081] y block =P(L(A(C(x))))

[0082] Where C represents the convolution operation, A represents the activation function, L represents the normalization operation, P represents the pooling operation, and y block It represents the feature vector output after being processed by the above network block. The recognition network block will be repeated m times in the network to enhance the feature extraction capability of the network.

[0083] Then, we use RNN and its variant networks in conjunction. By sending the vector results processed by the recognition network block to two LSTM layers respectively, we make them learn the characteristics of different aspects of the sequence independently. The formula is:

[0084] h 1 =LSTM 1 (y block ; Θ 1 )

[0085] h 2 =LSTM 2 (y block ; Θ 2 )

[0086] where y block is the feature vector output by the CNN network, Θ 1 and θ 2 Represent the parameters of the two LSTM networks, h 1 and h 2 They are the hidden states of the two LSTM networks at the last moment of the time series. Then use the concatenate function to stack the outputs of the two LSTM layers.

[0087] Repeat the dual LSTM network + Concatenate operation n times to effectively represent the multi-frame context content features x global_embedding .

[0088] The vector is then sent to the fully connected layer for feature integration and mapping.

[0089] Finally, three different fully connected networks (FCNs) are used to extract different potential feature vectors that distinguish models, states, and systems, namely

[0090] x type =φ type (W type x global_embedding )

[0091] x mode =φ mode (W mode x global_embedding )

[0092] x class =φ class (W intent x global_embedding )

[0093] In the formula, W represents the weight matrix of the fully connected layer, and φ represents the nonlinear activation function of the output layer. Finally, the three different potential representation vectors that distinguish model, state, and system are input into the corresponding classifiers to obtain the corresponding prediction results.

[0094]

[0095] Where W classfication is the weight matrix of the classifier.

[0096] Step S2, recognition model training: first construct the training data set required for training the intelligent recognition model, then train and generate the prior knowledge model required for model, state and system intelligent recognition, and finally store the model training hyperparameter results, the network architecture of the recognition model and its trained weight coefficients in the recognition knowledge file to prepare for subsequent model testing.

[0097] Furthermore, the step S2 includes:

[0098] Step S21, preprocessing the PDW sequence. First, sample the PDW sequence to reduce the amount of data, and the sampled PDW sequence is defined as the SDW sequence. Secondly, according to the timestamp in the SDW sequence data, take the data frame length T and the inter-frame interval ΔT, and perform frame interception on the SDW sequence. According to the timestamp in the SDW sequence data, the timestamp, sequence features, and sequence labels in the same frame of SDW data are divided and stored for subsequent training.

[0099] Step S22: construct the training data set D required for training the intelligent recognition model train For each annotated process data file in the training file path set, according to the agreed data frame length, randomly sample training sample pairs proportional to the total length of the process data to construct the training data set D train

[0100]

[0101] Where N tratn represents the number of process data files in the training set, M i represents the number of samples collected by the i-th process data file, Δ represents the duration of the sample frame, represents the jth sample collected in the i-th file, represents the label vector of the jth sample collected in the i-th file, Indicates that the jth sample collected in the i-th file contains a pulse sequence quantity,

[0102] Step S23: training and generating the prior knowledge model required for intelligent identification of model, state and system Where θ is the model parameter, θ hp is the hyperparameter of the model. Based on the multi-input multi-task learning neural network architecture, the multi-task optimization loss function L is set. train The sample pairs in are used as input, and the gradient descent training method is used to train the model parameters of the intelligent recognition model according to the set batch size, number of iterations and optimization strategy. The multi-task optimization loss function L is defined as

[0103] L=λ type L type +λ mode L mode +λ class L class ,∑λ i =1

[0104]

[0105] In the formula, λ type , mode and λ class are the weight coefficients for model type, working mode and working system identification classification, L type , L mode and L class are the loss functions for model type, working mode and working system identification and classification, respectively, C type , C mode and C class are the number of categories of model type, working mode and working system, and are the true labels of the i-th sample in the j-th category of model type, working mode and working system, respectively. and are the predicted probabilities of the i-th sample in the j-th category of model type, working mode and working system, respectively, and n is the number of training batches.

[0106] Step S24, first store the hyperparameters required for intelligent recognition preprocessing, data normalization, etc. in the specified recognition knowledge file, and then additionally store the network architecture of the aforementioned recognition model and the weight coefficients obtained by training in the specified recognition knowledge file.

[0107] Step S3, recognition model test: first establish the model evaluation index, then load the hyperparameters required by the model and the network architecture and parameters of the recognition model, then rely on the loaded intelligent recognition model to identify the data frame sequence in sequence, calculate the model recognition accuracy, and save the recognition test results for subsequent evaluation.

[0108] Furthermore, the step S3 includes:

[0109] Step S31, model evaluation index establishment:

[0110] Firstly, the model evaluation index is established. For each radar emitter tracking batch, K possible model type recognition results need to be given for each frame time f. Indicates that the model type estimated by batch number t in frame f is k, otherwise it is 0. t,f Representing the radiation source associated with the tracking lot number t, the true model type indicator function can be obtained. When the model type of the radiation source associated with the tracking batch number t in frame f is k, its value is 1, otherwise it is 0. With this concept, the confusion matrix can be used to measure the model type classification index of the electronic warfare system. The confusion matrix elements are different model types, and their values ​​are the number of true positives, true negatives, false positives and false negatives, that is,

[0111]

[0112] where f t and F t are the first and last frames of the tracking activity of batch number t, respectively. Based on the above confusion matrix, we define the recognition accuracy of model type k as

[0113]

[0114] Step S32, parameter model loading:

[0115] Load the required hyperparameters and identify the model The network architecture and its parameters are obtained by first loading the intelligent recognition hyperparameters from the specified recognition knowledge file to initialize the data preprocessing and data normalization processing functions. Then, the intelligent recognition model content is loaded, the neural network architecture of multi-input multi-task learning is reconstructed, and the network is initialized using the model parameters.

[0116] Step S33, specific test and result storage:

[0117] Relying on the loaded intelligent recognition model Identify the data frame sequence in turn, calculate the model recognition accuracy, and save the recognition test results. i Each sample frame in Input into the intelligent recognition model in sequence Then you can get the prediction results of model type, working mode and working system identification

[0118]

[0119] The predicted result and the true value label are paired The sequence is saved in the test result file corresponding to the i-th process data file for subsequent evaluation.

[0120] In other embodiments of the present invention, in order to clearly describe the implementation process and effectiveness of the method of the present invention, another specific implementation example is proposed herein.

[0121] The overall architecture block diagram of the radar emitter multi-task integrated identification method based on deep learning disclosed in this embodiment is as follows: Figure 1 As shown in the figure, the potential features of the original signal sequence of the frame are automatically extracted through the CNN network, RNN network and its variants. By capturing the distinguishing sequence features of different levels such as waveform style, resource scheduling and mode switching, the model, state and system identification of the radiation source are realized in one. The specific implementation steps are as follows:

[0122] The method of this embodiment specifically includes three processes: recognition model construction, model training and model reasoning:

[0123] Step 1, recognition model construction: construct an abstract recognition model. Specifically, in this example, three network layers, namely, one-dimensional convolution layer, batch normalization layer, and maximum pooling layer, constitute a recognition network block, which will be repeated three times in the network. The network operation of double LSTM+Concatenate is repeated twice.

[0124] Step 2: Identification model training: mainly includes two parts: data set preparation and model training.

[0125] Step 2-1, data set preparation: First, in order to improve the algorithm calculation rate and meet the real-time requirements, the PDW sequence data is sampled and slowed down to obtain SDW sequence data, where the SDW data contains six-dimensional parameters (sequence features) of TOA, PW, RF, PA, DOA and PRI, and the sequence labels contain 7 types of simulated radiation sources, two radar working modes (TWS and TAS), and two working systems (M-SCAN and E-SCAN). Specifically, in this embodiment, a total of 364 SDW sequence data are obtained by sampling the PDW sequence data obtained by simulation. According to the set data frame duration, the content of each process data file in the test file path set is discretized in a non-overlapping manner to form a discrete processing frame sequence, thereby obtaining SDW frame data N. train , and store the timestamp data time_frames, feature data features_frames, and label data labels_frames contained in the same frame sequence separately. Then, the frame sequence data are integrated to form a training data set. Specifically, in this implementation example, the frame length Δ=5s, the inter-frame interval ΔT=50ms, and the number of frames N in the discrete frame processing result trainIt is 150359. Among the feature data, the number of category feature vectors Type1 to Type7 are 309, 4719, 7315, 43473, 34527, 48953, and 11063 respectively; the number of mode feature vectors TWS and TAS are 66458 and 83901 respectively; the number of system feature vectors M-SCAN and E-SCAN are 7624 and 142735 respectively.

[0126] Step 2-2, model training: send the SDW frame data in the prepared data set to the neural network for training in batches, and set basic network parameters such as batch size, the ratio of the validation set to the total data set, and the total number of training rounds. Specifically, in this example, batch size is 1024, then the neural network input vector dimension is [1024,500,5], and val_ratio = 0.1, epoches = 100.

[0127] After the data is sent into the network, the initial dimension of the vector is [1024,500,5]. Then the one-dimensional convolution layer Conv1D is used to transform the dimension and expand the vector dimension so that the network can learn more complex feature representations.

[0128] The specific formula is:

[0129]

[0130] Among them, y t,c is the value of the output feature map at time step t, b c is the bias term, D is the number of channels of the input feature map, and w c,d,i Make the convolution kernel weights for output channel c and output channel d, i represents the position in the convolution kernel, -k to k represents the range covered by the convolution kernel, x t+i,d Let the value of the input feature map d at time step t+i, specifically, in this example, D=5 in the one-dimensional convolution layer, dim(w c,d,i )=128, convolution kernel size kener_sieze=5, convolution kernel moving step step=1, activation function f uses ReLU.

[0131] Afterwards, the one-dimensional convolution result is sent to the batch normalization layer to improve training speed and stability.

[0132] The specific formula is:

[0133]

[0134] Where x is the input data, μ B is the mean of the mini-batch data, Let the variance of the mini-batch data, ∈ be a very small constant for numerical stability. γ and β are learnable parameters corresponding to the scale and offset, respectively. Specifically, the batch size is 1024 in this example.

[0135] Afterwards, the vector is sent to the maximum pooling layer to reduce the vector dimension to reduce the amount of model calculation, while keeping the number of vector output channels unchanged to retain multi-dimensional feature information.

[0136] The specific formula is:

[0137]

[0138] Among them, P i,j is the value of the feature map at position (i, j) after pooling, The input feature map is in the window (k, l), and the value in the local area starting from (i, j) is obtained. k and l represent the index of the pooling window. Specifically, in this layer of the example, the pooling window size is 4, the pooling step is 2, and the padding method is valid (that is, no padding is performed on the edges). After the one-dimensional convolution layer and the batch normalization layer, the dimension of the vector becomes [1024, 500, 128]. After the maximum pooling layer, the dimension of the vector becomes [1024, 249, 128].

[0139] After that, the processing flow of the one-dimensional convolution layer, batch normalization layer, and maximum pooling layer is repeated twice to strengthen data feature extraction and abstraction, while achieving dimensionality reduction and reducing overfitting. Specifically, in this implementation example, in addition to the change in vector dimension in the subsequent two repeated layer processing, only the pooling step size of the maximum pooling layer is changed to 3 and 4 respectively, and the other parameters are not changed. After the above steps, the vector dimension changes from [1024, 500, 5] to [1024, 20, 32] after a total of three repetitions of the one-dimensional convolution layer, batch normalization layer, and maximum pooling layer processing flow.

[0140] After that, the vectors are sent to two LSTM layers respectively, so that they can learn different aspects of the sequence independently, and then the two LSTM layer outputs are stacked using the concatenate function. Specifically, in this example, the drop ratio of the LSTM layer is taken as 0.5, and the activation function is Sigmoid. After passing through the LSTM layer, the vector dimension [1024, 20, 36] remains unchanged, and becomes [1024, 20, 64] after stacking. Then repeat the double LSTM + concatenate stacking operation again, so that the vector dimension becomes [1024, 64].

[0141] After that, we enter the classification stage. First, the vector is sent to the fully connected layer for feature integration and mapping. The specific formula is:

[0142] y=f(W*x+b)

[0143] Among them, W is the weight matrix with the dimension (k, n), b is the bias vector, and f is the activation function. To prevent overfitting, the vector is sent to a Dropout layer, and the degree of regularization of the neural network is controlled by setting the dropout rate p. Specifically, in this example, the dimension of the fully connected layer output vector y is [1024, 128], the dimension of the weight matrix W is [128.64], the dimension of b is [128, 1], and the activation function is ReLU.

[0144] To prevent overfitting, the vector is sent to a Dropout layer, and the degree of regularization of the neural network is controlled by setting the dropout rate p. Specifically, in this example, p = 0.5.

[0145] Finally, the vector is sent to the final fully connected layer, and the vector output dimension is determined according to the number of classifications to be implemented to achieve classification. Specifically, in this embodiment, the output dimension of the mode classification vector is [1024, 2], the output dimension of the category classification vector is [1024, 7], and the output dimension of the system classification vector is [1024, 2].

[0146] At this point, the model training is completed, and the best parameter model is saved based on the quality of the training results.

[0147] Step 3, model testing: The main steps are establishing evaluation indicators and model testing.

[0148] Step 31, evaluation indicators are established:

[0149] The evaluation index is established according to the index establishment method in the invention content.

[0150] Step 32, model testing:

[0151] The performance of the intelligent recognition model is tested using test data. First, the model trained in step 2 is loaded, and then the selected test data set is loaded. In this example, the selected test data set is the same as the training data set. Specifically, in this implementation example, the number of frames in the discrete frame processing result in the test set is 150359. Then the model is used to test the test set data, and the model prediction accuracy is calculated based on the test results. Among them, Figure 3 This is a sequence recognition result of Type 4 radiation source. In this recognition, the pattern recognition accuracy is 96.68%, the type recognition accuracy is 98.10%, and the system recognition accuracy is 98.10%. Figure 4 This is the overall recognition accuracy result represented by the confusion matrix in this example experiment.

[0152] It can be seen from the experimental results that the accuracy of integrated radiation source identification achieved by the method of the present invention is not lower than that of the general radiation source single-task identification method, and the integrated identification of radiation source type, working mode and working system is successfully realized. The method is innovative and advanced.

[0153] It should be noted that within the scope of protection defined in the claims of the present invention, the following embodiments can be combined and / or expanded or replaced in any logical way from the above specific implementation methods, such as disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0154] Example 1

[0155] A multi-task integrated identification method for radar emitters based on deep learning, comprising the following steps:

[0156] S1, identification model construction: by defining the characteristic space of the radar radiation source, the type, working mode and working system of the radar radiation source are characterized to realize the identification model construction;

[0157] S2, recognition model training: construct a training data set, and then use the recognition model constructed in step S1 to train the prior knowledge model, and store the model training hyperparameter results, the network architecture of the recognition model and its trained weight coefficients in the recognition knowledge file, in preparation for the model test in step S3;

[0158] S3, recognition model test: load the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters, then use the test data to test the performance of the intelligent recognition model, and finally rely on the loaded intelligent recognition model to identify the data frame sequence in sequence, and save the recognition test results for subsequent evaluation.

[0159] Example 2

[0160] On the basis of Example 1, in step S1, the feature space of the radar radiation source is defined, and the radar radiation source model type, working mode, and working system are characterized to realize the construction of the recognition model, which specifically includes the following sub-steps:

[0161] Step S11, firstly divide the signal pulse sequence of each tracking batch number in a non-overlapping manner according to a certain time length Δ Forming discrete processing frames

[0162]

[0163] Where F represents the signal pulse sequence of the radiation source tracking batch number i The number of discrete frames that can be divided; for each frame pulse sequence

[0164] Step S12, using a deep neural network to further extract the global sequence potential features x between multiple frames global_embedding The deep neural network includes convolutional neural network CNN, recurrent neural network RNN ​​and its variant networks; the specific network construction is as follows:

[0165] The CNN network uses a one-dimensional convolutional layer Conv1D to transform the dimension and expand the vector dimension, so that the network can learn more complex feature representations. The formula is as follows:

[0166] y CNN =CNN(x;θ)

[0167] Among them, x represents the input signal sequence frame, θ represents the parameters of the network, and y CNN Represents the feature vector output after processing by the CNN network;

[0168] After that, the one-dimensional convolution result is sent to the batch normalization layer to improve the training speed and stability; the variable is then sent to the maximum pooling layer to reduce the vector dimension to reduce the model calculation amount;

[0169] The above three networks constitute the identification network block, which is expressed as:

[0170] y block =P(L(A(C(x))))

[0171] Among them, C represents the convolution operation, A represents the activation function, L represents the normalization operation, P represents the pooling operation, and y block represents the feature vector output after being processed by the network block; the recognition network block will be repeated m times in the network to enhance the feature extraction capability of the network;

[0172] After that, we use RNN and its variant networks in conjunction. By sending the vector results processed by the recognition network block to two LSTM layers respectively, we make them learn the characteristics of different aspects of the sequence independently. The formula is:

[0173] h 1 =LSTM 1 (y block ; Θ 1 )

[0174] h 2 =LSTM 2 (y block ; Θ 2 )

[0175] Among them, y block is the feature vector output by the CNN network, Θ 1 and θ2 Represent the parameters of the two LSTM networks, h 1 and h 2 are the hidden states of the two LSTM networks at the last moment of the time series; the concatenate function is used to stack the outputs of the two LSTM layers; the dual LSTM network and Concatenate operations are repeated n times to effectively represent the multi-frame context content features x global_embedding ;

[0176] Afterwards, the vector is sent to the fully connected layer for feature integration and mapping;

[0177] Finally, three different fully connected networks (FCNs) are used to extract different potential feature vectors that tend to distinguish models, states, and systems, namely:

[0178] x type =φ type (W tupe x global_embedding )

[0179] x mode =φ mode (W mode x global_embedding )

[0180] x class =φ class (W intent x global_embedding )

[0181] In the formula, W represents the weight matrix of the fully connected layer, and φ represents the nonlinear activation function of the output layer. Finally, the three different potential representation vectors that distinguish the model, state, and system are input into the corresponding classifiers to obtain the corresponding prediction results:

[0182]

[0183]

[0184] Where W classfication is the weight matrix of the classifier.

[0185] Example 3

[0186] On the basis of Example 1, in step S2, the training data set is constructed, and the recognition model constructed in step S1 is used to train the prior knowledge model, and the model training hyperparameter results, the network architecture of the recognition model and the weight coefficients obtained by training are stored in the recognition knowledge file, in preparation for the model test in step S3, which specifically includes the following sub-steps:

[0187] S21, preprocessing the PDW sequence: first, sampling the PDW sequence to reduce the amount of data, and the sampled PDW sequence is defined as the SDW sequence; secondly, according to the timestamp in the SDW sequence data, the data frame length T and the inter-frame interval ΔT are taken to perform frame interception on the SDW sequence; according to the timestamp in the SDW sequence data, the timestamp, sequence features, and sequence labels in the same frame of SDW data are divided and stored for subsequent training;

[0188] Step S22: construct the training data set D required for training the intelligent recognition model train ; For each annotated process data file in the training file path set, according to the agreed data frame length, randomly sample training sample pairs proportional to the total length of the process data to construct the training data set D train ;

[0189]

[0190] Where N train represents the number of process data files in the training set, M i represents the number of samples collected by the i-th process data file, Δ represents the duration of the sample frame, represents the jth sample collected in the i-th file, represents the label vector of the jth sample collected in the i-th file, Indicates that the jth sample collected in the i-th file contains a pulse sequence quantity;

[0191] Step S23: training and generating the prior knowledge model required for intelligent identification of model, state and system Where θ is the model parameter, θ hp is the hyperparameter of the model; according to the multi-input multi-task learning neural network architecture built, and the multi-task optimization loss function L is set; with the training data set D train The sample pairs in are used as input to train the model parameters of the recognition model; the multi-task optimization loss function L is defined as:

[0192] L=λ type L type +λ mode L mode +λ class L class ,∑λ i =1

[0193]

[0194] In the formula, λ type , mode and λ classare the weight coefficients for model type, working mode and working system identification classification, L type , L mode and L class are the loss functions for model type, working mode and working system identification and classification, respectively, C type , C mode and C class are the number of categories of model type, working mode and working system, and are the true labels of the i-th sample in the j-th category of model type, working mode and working system, respectively. and are the predicted probabilities of the i-th sample in the j-th category of model type, working mode and working system, respectively, and n is the number of batches for training;

[0195] Step S24, firstly store the hyperparameters in the designated recognition knowledge file, and then additionally store the network architecture of the aforementioned recognition model and the weight coefficients obtained through training in the designated recognition knowledge file.

[0196] Example 4

[0197] On the basis of Example 3, in step S23, the model parameters of the training recognition model are specifically trained using a gradient descent training method according to a set batch size, number of iterations and optimization strategy.

[0198] Example 5

[0199] On the basis of Example 3, in step S24, the hyperparameters are specifically hyperparameters required for preprocessing of intelligent recognition and data normalization.

[0200] Example 6

[0201] On the basis of Example 1, in step S3, the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters are loaded, and the performance of the intelligent recognition model is tested using the test data. Finally, the data frame sequence is sequentially recognized based on the loaded intelligent recognition model, and the recognition test results are saved for subsequent evaluation, which specifically includes the following sub-steps:

[0202] S31, model evaluation index establishment: firstly, the model evaluation index is established, and K possible model type recognition results are required to be given for each frame time f within the duration of each radar radiation source tracking batch; Indicates that the model type estimated by batch number t in frame f is k, otherwise it is 0; record a t,f Represents the radiation source associated with the tracking lot number t, and obtains the true model type indicator function When the model type of the radiation source associated with the tracking lot number t in frame f is k, its value is 1, otherwise it is 0;

[0203] The confusion matrix is ​​used to quantify the model type classification index of the electronic warfare system; the confusion matrix elements are different model types, and their values ​​are the number of true positives, true negatives, false positives and false negatives, that is,

[0204]

[0205] Among them, f t and F t are the first and last frames of the tracking activity of batch number t respectively; according to the above confusion matrix, the recognition accuracy of model type k is defined as

[0206]

[0207] Step S32, parameter model loading: load the required hyperparameters and recognition model The network architecture and its parameters are obtained; first, the intelligent recognition hyperparameters are loaded from the specified recognition knowledge file to initialize the data preprocessing and data normalization processing functions; then, the intelligent recognition model content is loaded, the neural network architecture of multi-input multi-task learning is reconstructed, and the network is initialized using the model parameters;

[0208] Step S33, specific test and result saving: relying on the loaded intelligent recognition model Identify the data frame sequence in turn, calculate the model recognition accuracy, and save the recognition test results; each discrete data frame sequence in the test set o i Each sample frame in Input into the intelligent recognition model in sequence Get the prediction results of model type, working mode and working system identification

[0209]

[0210] The predicted result and the true value label are paired The sequence is saved in the test result file corresponding to the i-th process data file for subsequent evaluation.

[0211] Example 7

[0212] A multi-task integrated identification device for radar radiation sources based on deep learning, comprising a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method described in any one of Embodiments 1 to 6 is executed.

[0213] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.

[0214] According to one aspect of an embodiment of the present invention, a computer program product or a computer program is provided, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations.

[0215] As another aspect, an embodiment of the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.

Claims

1. A multi-task integrated identification method for radar emitters based on deep learning, characterized in that: The following steps are involved: S1, identification model construction: by defining the characteristic space of the radar radiation source, the type, working mode and working system of the radar radiation source are characterized to realize the identification model construction; S2, recognition model training: construct a training data set, and then use the recognition model constructed in step S1 to train the prior knowledge model, and store the model training hyperparameter results, the network architecture of the recognition model and its trained weight coefficients in the recognition knowledge file, in preparation for the model test in step S3; S3, recognition model test: load the hyperparameters stored in step S2, the network architecture of the recognition model and its parameters, then use the test data to test the performance of the intelligent recognition model, and finally rely on the loaded intelligent recognition model to identify the data frame sequence in sequence, and save the recognition test results for subsequent evaluation.

2. The multi-task integrated identification method of radar emitters based on deep learning according to claim 1 is characterized in that: In step S1, the feature space of the radar radiation source is defined, and the radar radiation source model type, working mode, and working system are characterized to realize the construction of the recognition model, which specifically includes the following sub-steps: Step S11, firstly divide the signal pulse sequence of each tracking batch number in a non-overlapping manner according to a certain time length Δ Forming discrete processing frames Where F represents the signal pulse sequence of the radiation source tracking batch number i The number of discrete frames that can be divided; for each frame pulse sequence Step S12, using a deep neural network to further extract the global sequence potential features x between multiple frames global_embedding The deep neural network includes convolutional neural network CNN, recurrent neural network RNN ​​and its variant networks; the specific network construction is as follows: The CNN network uses a one-dimensional convolutional layer Conv1D to transform the dimension and expand the vector dimension, so that the network can learn more complex feature representations. The formula is as follows: y CNN =CNN(x;θ) Among them, x represents the input signal sequence frame, θ represents the parameters of the network, and y CNN Represents the feature vector output after processing by the CNN network; After that, the one-dimensional convolution result is sent to the batch normalization layer to improve the training speed and stability; the variable is then sent to the maximum pooling layer to reduce the vector dimension to reduce the model calculation amount; The above three networks constitute the identification network block, which is expressed as: y block =P(L(A(C(x)))) Among them, C represents the convolution operation, A represents the activation function, L represents the normalization operation, P represents the pooling operation, and y block represents the feature vector output after being processed by the network block; the recognition network block will be repeated m times in the network to enhance the feature extraction capability of the network; After that, we use RNN and its variant networks in conjunction. By sending the vector results processed by the recognition network block to two LSTM layers respectively, we make them learn the characteristics of different aspects of the sequence independently. The formula is: h1=LSTM1(y block ;Θ1) h2=LSTM2(y block ;Θ2) Among them, y block is the feature vector output by the CNN network, Θ1 and Θ2 represent the parameters of the two LSTM networks, h1 and h2 are the hidden states of the two LSTM networks at the last moment of the time series; the concatenate function is then used to stack the outputs of the two LSTM layers; the dual LSTM network and Concatenate operations are repeated n times to effectively characterize the multi-frame context content features x global_embedding ; Afterwards, the vector is sent to the fully connected layer for feature integration and mapping; Finally, three different fully connected networks (FCNs) are used to extract different potential feature vectors that tend to distinguish models, states, and systems, namely: x type =φ type (E type x global_embedding ) x mode =φ mode (W mode x global_embedding ) x class =φ class (W intent x global_embedding ) In the formula, W represents the weight matrix of the fully connected layer, and φ represents the nonlinear activation function of the output layer. Finally, the three different potential representation vectors that distinguish the model, state, and system are input into the corresponding classifiers to obtain the corresponding prediction results: Where W classfication is the weight matrix of the classifier.

3. The multi-task integrated identification method of radar emitters based on deep learning according to claim 1 is characterized in that: In step S2, the training data set is constructed, and the recognition model constructed in step S1 is used to train the prior knowledge model, and the model training hyperparameter results, the network architecture of the recognition model and the weight coefficients obtained by training are stored in the recognition knowledge file, in preparation for the model test in step S3, which specifically includes the following sub-steps: S21, preprocessing the PDW sequence: first, sampling the PDW sequence to reduce the amount of data, and the sampled PDW sequence is defined as the SDW sequence; secondly, according to the timestamp in the SDW sequence data, the data frame length T and the inter-frame interval ΔT are taken to perform frame interception on the SDW sequence; according to the timestamp in the SDW sequence data, the timestamp, sequence features, and sequence labels in the same frame of SDW data are divided and stored for subsequent training; Step S22: construct the training data set D required for training the intelligent recognition model train ;against For each annotated process data file in the training file path set, according to the agreed data frame duration, randomly sample training sample pairs proportional to the total length of the process data to construct the training data set D train ; Where N train represents the number of process data files in the training set, M i represents the number of samples collected by the i-th process data file, Δ represents the duration of the sample frame, represents the jth sample collected in the i-th file, represents the label vector of the jth sample collected in the i-th file, Indicates that the jth sample collected in the i-th file contains a pulse sequence quantity; Step S23: training and generating the prior knowledge model required for intelligent identification of model, state and system Where θ is the model parameter, θ hp is the hyperparameter of the model; according to the multi-input multi-task learning neural network architecture built, and the multi-task optimization loss function L is set; with the training data set D train The sample pairs in are used as input to train the model parameters of the recognition model; the multi-task optimization loss function L is defined as: L=λ type L type +λ mode L mode +λ class L class ,∑λ i =1 In the formula, λ type , mode and λ class are the weight coefficients for model type, working mode and working system identification classification, L type , L mode and L class are the loss functions for model type, working mode and working system identification and classification, respectively, C type , C mode and C class are the number of categories of model type, working mode and working system, and are the true labels of the i-th sample in the j-th category of model type, working mode and working system, respectively. and are the predicted probabilities of the i-th sample in the j-th category of model type, working mode and working system, respectively, and n is the number of batches for training; Step S24, firstly store the hyperparameters in the designated recognition knowledge file, and then additionally store the network architecture of the aforementioned recognition model and the weight coefficients obtained through training in the designated recognition knowledge file.

4. The multi-task integrated identification method of radar emitters based on deep learning according to claim 3 is characterized in that: In step S23, the model parameters of the training recognition model are specifically trained using a gradient descent training method according to a set batch size, number of iterations and optimization strategy.

5. The multi-task integrated identification method of radar emitters based on deep learning according to claim 3 is characterized in that: In step S24, the hyperparameters are specifically hyperparameters required for preprocessing and data normalization of intelligent recognition.

6. The multi-task integrated identification method of radar emitters based on deep learning according to claim 1 is characterized in that: In step S3, the hyperparameters, network architecture and parameters of the recognition model stored in step S2 are loaded, and the performance of the intelligent recognition model is tested using the test data. Finally, the data frame sequence is sequentially recognized based on the loaded intelligent recognition model, and the recognition test results are saved for subsequent evaluation, which specifically includes the following sub-steps: S31, model evaluation index establishment: firstly, the model evaluation index is established, and K possible model type recognition results are required to be given for each frame time f within the duration of each radar radiation source tracking batch; Indicates that the model type estimated by batch number t in frame f is k, otherwise it is 0; record a t,f Represents the radiation source associated with the tracking lot number t, and obtains the true model type indicator function When the model type of the radiation source associated with the tracking lot number t in frame f is k, its value is 1, otherwise it is 0; The confusion matrix is ​​used to quantify the model type classification index of the electronic warfare system; the confusion matrix elements are different model types, and their values ​​are the number of true positives, true negatives, false positives and false negatives, that is, Among them, f t and F t are the first and last frames of the tracking activity of batch number t respectively; according to the above confusion matrix, the recognition accuracy of model type k is defined as Step S32, parameter model loading: load the required hyperparameters and recognition model The network architecture and its parameters; first load the intelligent recognition hyperparameters from the specified recognition knowledge file to initialize the data preprocessing and data normalization processing functions; then load the intelligent recognition model content, reconstruct the neural network architecture of multi-input multi-task learning, and use the model parameters to initialize the network; Step S33, specific test and result saving: relying on the loaded intelligent recognition model Identify the data frame sequence in turn, calculate the model recognition accuracy, and save the recognition test results; each discrete data frame sequence in the test set o i Each sample frame in Input into the intelligent recognition model in sequence Get the prediction results of model type, working mode and working system identification The predicted result and the true value label are paired The sequence is saved in the test result file corresponding to the i-th process data file for subsequent evaluation.

7. A multi-task integrated identification device for radar radiation sources based on deep learning, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method according to any one of claims 1 to 6 is executed.

Images