A radar jamming recognition method based on LSTM and variational autoencoder

Abstract

The application belongs to the technical field of radar jamming recognition, and discloses a radar jamming recognition method based on LSTM and a variational autoencoder. The application encodes time series signals by constructing a joint network of LSTM and a variational autoencoder to obtain corresponding feature vectors, then splices the feature vectors of the transmitted signals and echo signals to obtain a combination vector capable of representing the jamming type, and then classifies and learns the combination vector through a convolutional neural network to realize the recognition of the jamming signals. The application can effectively process time series data and learn the complex distribution of data by virtue of the ability of the LSTM network to solve long-term dependence and the advantages of the variational autoencoder in dimension reduction and feature extraction; meanwhile, the splicing of the vectors before and after the jamming process to represent the change relationship of the signals before and after the jamming process is helpful to adapt to the recognition of jamming patterns under any change.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the technical field of radar jamming recognition, and particularly relates to a radar jamming recognition method based on LSTM and a variational autoencoder. BACKGROUND

[0002] Electromagnetic interference has a serious impact on radar detection, and selecting appropriate waveforms, signal processing methods, etc. based on interference patterns is a common means of resisting electromagnetic interference. The performance of interference pattern recognition plays an important role in subsequent anti-jamming, and therefore, interference pattern recognition has attracted widespread attention.

[0003] In interference recognition related research, one commonly used method is to obtain a time-frequency image by performing time-frequency analysis on the received signal, and to convert the interference recognition problem into an image recognition problem, but this method is not time-efficient, and its performance largely depends on the pre-processing effect of the time-frequency image, including frequency resolution, start and end time jumps, etc. Another commonly used method is to directly process the time sequence signal, such as directly classifying and recognizing the original time domain sequence of the interference signal, but the recognition performance is poor under low jamming noise ratio.

[0004] In recent years, with the rise of machine learning and deep learning, interference recognition technology has also entered an intelligent development stage. After feature extraction, methods such as decision tree, back propagation neural network (BP) and support vector machine (SVM) can be used to recognize the type of active jamming signal, and good performance has been achieved. At the same time, the use of convolutional neural network (CNN) model can also achieve recognition of radar jamming signals. However, these methods usually assume that the radar transmitted waveform is fixed, i.e. the radar jamming signal recognition network is implemented under a certain specific transmitted waveform. For cognitive radar, the radar transmitted signal will change with the transmission, and the interference pattern will also change with the transmitted waveform. Due to the large parameter space of the transmitted waveform and the numerous interference patterns, the parameter space of the interference signal is huge, making it difficult for conventional interference recognition networks to adapt to interference pattern recognition under arbitrary changes, i.e. not robust and universal for multiple transmitted signals. SUMMARY

[0005] The present application aims to provide a radar jamming recognition method based on LSTM and a variational autoencoder, which can recognize the interference type when the transmitted signal changes, without being limited to changes in the transmitted signal.

[0006] To achieve the above-mentioned purpose, the present application provides a radar jamming recognition method based on LSTM and a variational autoencoder, comprising the following steps:

[0007] Step 1: training a feature extraction network

[0008] A feature extraction network is created using a long short-term memory network (LSTM) and a variational autoencoder (VAE) network, which includes an encoder, a fully connected layer, and a decoder.

[0009] The encoder uses an LSTM to process the input sequence, extracting the hidden state of the feature extraction network as a feature vector. The input sequence includes the transmitted signal s(t) and the disturbed echo signal j(t). The extracted feature vector is the transmitted feature vector and the interference feature vector.

[0010] The fully connected layer includes a linear layer, which constructs a latent variable based on the parameter distribution of the feature extraction network hidden state.

[0011] The decoder uses an LSTM to decode, decoding the latent variable into a reconstructed output sequence.

[0012] Step 2: training a classifier

[0013] A classifier is created based on a convolutional neural network, which includes one or more convolutional modules composed of a two-dimensional convolution kernel, a BN layer, an activation function, and a pooling layer. The bottom layer of the classifier is a fully connected layer.

[0014] The softmax function is used as the classifier output interference pattern classification result.

[0015] Step 3: input the transmitted signal s(t) and the disturbed echo signal j(t) to be identified, and use the trained feature extraction network and classifier to identify the interference pattern:

[0016] S1: encode the transmitted signal s(t) and the received disturbed echo signal j(t) respectively, and obtain the hidden state of the encoded feature extraction network as the feature vector of the signal. The obtained feature vector is the transmitted feature vector and the interference feature vector.

[0017] S2: concatenate the two feature vectors in step S1 to obtain a concatenated vector y that can represent the relationship between the transmitted signal s(t) and the disturbed echo signal j(t).

[0018] S3: perform a one-dimensional convolution operation on the concatenated vector y to transform the dimensions and obtain a tensor Y.

[0019] S4: use the classifier to perform classification learning on the transformed tensor Y, and output the interference pattern classification result.

[0020] Further, the input data size of the convolution module is (128, 1, 256, 1024), 128 represents 128 input sequences per batch; the channel number 1 represents a single channel, 256 is the dimension after dimension transformation, and the vector length is 1024.

[0021] Further, the loss function in the classifier selects a cross-entropy loss function, assuming that c i is the category of the radar jamming signal, p(c i ) is the target distribution, and q(c i ) is the predicted distribution. The cross-entropy of p and q is expressed as

[0022] Further, the parameters of the hidden state of the feature extraction network in step 1 are the mean and log variance of the emission feature vector and the jamming feature vector.

[0023] Further, the training method for training the feature extraction network in step 1 is centralized training:

[0024] S5: Combine the emission signal s(t) and the jammed echo signal j(t) of different styles to construct an input sequence;

[0025] S6: The input sequence is encoded by the long short-term memory network LSTM encoder to obtain the hidden state, and the mean and log variance are extracted from the hidden state. The latent variable is generated by the mean and log variance through reparameterization;

[0026] S7: The latent variable is decoded into a reconstructed output sequence by the long short-term memory network LSTM decoder;

[0027] S8: Use all signal loss means between the reconstructed output sequence and the input sequence as the total loss function to train the entire feature extraction network.

[0028] Further, the training method for training the feature extraction network in step 1 is separate training:

[0029] S9: Combine the emission signal s(t) and the jammed echo signal j(t) of each style as separate input sequences;

[0030] S10: The input sequence is encoded by the long short-term memory network LSTM encoder to obtain the hidden state, and the mean and log variance are extracted from the hidden state. The latent variable is generated by the mean and log variance through reparameterization;

[0031] S11: The latent variable is decoded into a reconstructed output sequence by the long short-term memory network LSTM decoder;

[0032] S12: Train the feature extraction network of the corresponding signal using the mean of signal loss between the reconstructed output sequence and the input sequence in each combination as the corresponding loss function.

[0033] Further, the training process of the feature extraction network comprises:

[0034] S13: Let the input sequence be x(t), and the hidden state of the feature extraction network h_enc obtained by passing the long short-term memory network through the LSTM encoder, h_enc = Encoder(x(t)), where Encoder() represents the encoder function;

[0035] S14: Extract the mean μ and the logarithmic variance logσ of the hidden state of the feature extraction network by using the linear layer μ, logσ 2 = ExtractMeanAndLogVar(h_enc) 2 ;

[0036] S15: Generate the latent variable z from the mean μ and the logarithmic variance logσ 2 by reparameterization, z = μ + noise * σ, where noise is a random noise sampled from the standard normal distribution, and σ represents the standard deviation;

[0037] S16: Decode the latent variable into the reconstructed output sequence x'(t), x'(t) = Decoder(z), Decoder() represents the decoder function.

[0038] Further, the tensor Y is Y = conv(y), and conv() represents the convolution function.

[0039] Further, the loss function loss of the feature extraction network includes the mean square error loss MSE(x, x') between the reconstructed output sequence and the input sequence of the feature extraction network and the KL divergence loss, and the KL divergence loss is The loss function is

[0040] loss = MSE(x, x') + kld_weight * KL(N(μ, σ2), N(0, 1)).

[0041] Beneficial effects:

[0042] The application provides a radar jamming recognition method based on LSTM and a variational autoencoder, and aims at the problem that a conventional radar jamming signal recognition network cannot adapt to jamming pattern recognition under any change. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a graph of a received signal under an SMP jamming pattern related to an embodiment of the application;

[0044] Figure 2 is a graph of a received signal under an SRJ jamming pattern related to an embodiment of the application;

[0045] Figure 3 is a flowchart of a radar jamming recognition method based on LSTM and a variational autoencoder related to an embodiment of the application;

[0046] Figure 4 is a feature extraction network model schematic diagram;

[0047] Figure 5 is an identification effect diagram of five jamming patterns related to an embodiment of the application;

[0048] Figure 6 is a recognition accuracy comparison diagram of a conventional method related to an embodiment of the application.

[0049] BRIEF DESCRIPTION OF DRAWINGS

[0050] x(t) is an input sequence; h_enc is a hidden state of a feature extraction network, h_enc = Encoder(x(t)); Encoder() is an encoder function; μ is a mean of a variational autoencoder latent distribution; logσ 2logvar; ExtractMeanAndLogVar() is an operation function for extracting mean and logvar; z is a latent variable; noise is a random noise sampled from a standard normal distribution; σ is a standard deviation of a latent distribution of a variational autoencoder network; x'(t) is a reconstructed output sequence: Decoder() is a decoder function; s(t) is a transmit signal; j(t) is a jammed echo signal; y is a concatenated vector; Y is a tensor; conv() is a convolution function; MSE(x, x') is a mean square error loss between a reconstructed output sequence of a feature extraction network and an input sequence; KL divergence loss is a KL(N(μ, σ 2 ), N(0, 1)); kld_weight is a KL divergence loss weight; loss is a loss function of the feature extraction network; c i is a category of a radar jamming signal; p(c i ) is a target distribution; q(c i ) is a predicted distribution. DETAILED DESCRIPTION

[0051] The preferred mechanisms and implemented methods of the present application are further described below in conjunction with the accompanying drawings and detailed description.

[0052] With the continuous development of the DRFM (Digital Radio Frequency Memory) repeater, active main lobe jamming has gradually become the most mainstream jamming method in electronic countermeasures. In the embodiments of the present application, several main lobe jamming patterns are identified, but it should be understood that this is only an example and not a limitation of the present application. The method of the present application is generally applicable to the identification of other radar jamming patterns.

[0053] In the conventional radar transceiver and signal processing process, when the recognition network perceives that the echo signal has been jammed, a game theory method is used to find out whether there is a suitable active or passive anti-jamming processing method. If the active anti-jamming processing method can be used to suppress, the transmit waveform does not need to be modified. If the active anti-jamming processing method cannot be used to suppress, a specific transmit waveform needs to be designed through the active anti-main lobe jamming method, so as to modify the transmit waveform and re-enter the processing process.

[0054] It can be seen that the radar transmit signal will change, and the jamming pattern will also change with the transmit signal. Then the relationship between the jamming signal and the transmit signal can be expressed as:

[0055] j(t) = f[s(t)] (Formula 1)

[0056] Where s(t) is the transmit signal, h(T) is the jammed echo signal, and f represents some relationship between the two.

[0057] And the radar received signal is composed of the transmitted signal s(t), the jammed echo signal j(t) and the noise signal n(t), that is

[0058] r(t) = s(t) + j(t) + n(t) (Formula 2)

[0059] Combined with (Formula 1), it can be seen that when the transmitted signal changes, the received signal of the radar will change. As shown in Figure 1 And Figure 2 When the pulse width and repetition frequency of the transmitted signal change, the existing spectrum dispersion jamming (SMSP) cannot effectively cover all change modes, at this time, the jamming device can select to sample specific pulse sequences and intermittently retransmit, that is, from spectrum dispersion jamming (SMSP) to intermittent sampling and repeated jamming (ISRJ), in order to improve the pertinence and efficiency of jamming. At the same time, combined with (Formula 2), when the transmitted signal and the jammed echo signal change, the received signal will also change, such as the waveform of the signal will change from continuous to intermittent, which makes it difficult for the recognition network to identify the changed jamming pattern.

[0060] In view of the problem that the above-mentioned jamming pattern changes with the change of the transmitted signal, which makes it difficult for the recognition network to identify the jamming pattern, as shown in Figures 1 to 6 The embodiment of the application discloses a radar jamming recognition method based on LSTM and variational auto-encoder. Figure 3 As shown in the figure, the embodiment of the application relates to a radar jamming recognition method based on LSTM and variational auto-encoder. The application combines a long short-term memory network (Long Short-Term Memory, LSTM) and a variational auto-encoder (Variational Auto-Encoder, VAE) network framework to construct a joint LSTM+VAE network to encode the input time series signal to obtain a feature vector, and finally construct a convolutional neural network as a classifier to complete the recognition of the feature vector.

[0061] Embodiment 1

[0062] A radar jamming recognition method based on LSTM and variational auto-encoder, characterized in that it comprises the following steps:

[0063] Step 1: training a feature extraction network

[0064] A feature extraction network is created by using a long short-term memory network LSTM and a variational auto-encoder VAE network. The feature extraction network combines the sequence processing ability of LSTM and the feature of the generative model of VAE, so that it can effectively process time series data, learn the complex distribution of data, and generate new data instances from the learned latent space. The processing of the feature extraction network model is as shown in Figure 4The feature extraction network includes an encoder, a fully connected layer and a decoder by means of the encoder, latent space and decoder framework of the VAE network, combined with the sequence processing capability of the LSTM;

[0065] The encoder uses a long short-term memory network (LSTM) to process the input sequence, extracts the hidden state of the feature extraction network as a feature vector, and the input sequence includes the transmitted signal s(t) and the echo signal j(t) disturbed by interference, and the extracted feature vector is a transmitted feature vector and an interference feature vector;

[0066] The fully connected layer includes a linear layer for implementing linear transformation, and the formula is: Y = XW + b, where X is an input matrix, W is a weight matrix, b is a bias vector, and Y is an output matrix. The specific code implementation is nn.Linear in PyTorch.

[0067] The latent variable is constructed according to the parameter distribution of the hidden state of the feature extraction network, that is, Figure 4 The sample vector in the formula is

[0068] The decoder uses a long short-term memory network (LSTM) to decode, and decodes the latent variable into a reconstructed output sequence.

[0069] Step 2: Train the classifier:

[0070] The classifier is created based on a convolutional neural network, and the classifier includes one or more convolutional modules composed of a two-dimensional convolution kernel, a BN layer, an activation function and a pooling layer, and the bottom layer of the classifier is a fully connected layer; The classifier in this embodiment includes three or more convolutional modules composed of a two-dimensional convolution kernel, a BN layer, an activation function and a pooling layer, and the bottom layer of the classifier is a fully connected layer.

[0071] The fully connected layer is located at the bottom of the convolutional neural network, and is used to map the feature space calculated by the previous layer to the label space of the sample, integrate the feature vectors together, and facilitate the final classification.

[0072] The softmax function is used as the classifier output interference pattern classification result.

[0073] The softmax function is a normalized exponential function, which is an activation function in a classification task, has a special mathematical expression, is used to convert a vector into a probability distribution, and the sum of all elements is 1, and the number of elements is the number of classifications.

[0074] Step 3: Input the transmitted signal s(t) and the echo signal j(t) disturbed by interference to be identified, and use the trained feature extraction network and classifier to identify the interference pattern:

[0075] S1: encode the transmitted signal s(t) and the received interference echo signal j(t) respectively, obtain the hidden state of the encoded feature extraction network as the feature vector of the signal, and the obtained feature vector is the transmitted feature vector and the interference feature vector;

[0076] S2: splice the two feature vectors in step S1 to obtain a spliced vector y capable of representing the relationship between the transmitted signal s(t) and the interference echo signal j(t);

[0077] S3: perform dimension transformation on the spliced vector y through one-dimensional convolution operation to obtain a tensor Y;

[0078] S4: using a classifier to classify and learn the transformed tensor Y, and output the interference pattern classification result.

[0079] In the training process of the model, the whole LSTM+VAE model is trained, and in the feature extraction, the encoder part of the model is used to extract the hidden state as the feature vector of the input sequence. The present application provides two training methods: centralized training and separate training, and the two processes are the same, only the data division is different. The sample transmitted signal and the echo signal interfered by different patterns form a first training data set, which is used to train the LSTM+VAE feature extraction network.

[0080] In the centralized training mode, the sample transmitted signal and the echo signal interfered by different patterns are combined to form an input sequence, the input sequence is input into the LSTM encoder to obtain the hidden state, the mean and the logarithmic variance are extracted from the hidden state, the latent variable is generated by reparameterization from the mean and the logarithmic variance, and the output sequence is reconstructed from the latent variable by the LSTM decoder; the total loss function is obtained by using the mean of all signal losses between the reconstructed output sequence and the input sequence, and the whole feature extraction network is trained.

[0081] In the separate training mode, the sample transmitted signal and the echo signal interfered by different patterns are respectively used as separate input sequences, wherein each interference echo signal forms an input sequence, and these input sequences are respectively input into the LSTM encoder to obtain the hidden state, the mean and the logarithmic variance are extracted from the hidden state, the latent variable is generated by reparameterization from the mean and the logarithmic variance, and the output sequence is reconstructed from the latent variable by the LSTM decoder; the mean of the loss between the reconstructed output sequence and the input sequence of each signal is used as the corresponding loss function, and the feature extraction network of the corresponding signal is trained.

[0082] Specifically, let the input sequence be x(t), and the hidden state h_enc be obtained through the LSTM encoder, i.e., h_enc = Encoder(x(t)), where Encoder() represents the encoder function. The encoder is constructed using an LSTM network, and the specific parameters are shown in Table 1:

[0083] Table 1 LSTM encoder network parameters

[0084] input_size hidden_size latent_size num_layers 12000 512 256 1

[0085] where input_size is the input sequence size, hidden_size is the hidden layer size, latent_size is the latent space size, and num_layers is the number of network layers.

[0086] Subsequently, the mean μ and the logarithmic variance logσ are extracted from the hidden state 2 i.e.,

[0087] μ, logσ 2 = ExtractMeanAndLogVar(h_enc) (Equation 3)

[0088] These two values are used to parameterize the distribution in the latent space. The latent variable z is generated by the mean μ and the standard deviation σ through reparameterization, i.e., z = μ + noise * σ. ExtractMeanAndLogVar() in Equation (8) is the operation of extracting the mean and the logarithmic variance, which is implemented by a linear layer (fully connected layer). The specific implementation is to perform matrix multiplication on the input data and add a bias vector. noise is a random noise sampled from a standard normal distribution.

[0089] Next, the output sequence is reconstructed from the latent variable through the decoder. That is

[0090] x'(t) = Decoder(z) (Equation 4) The decoder function Decoder() is also constructed by an LSTM network, and the network parameters are shown in Table 2:

[0091] Table 2 LSTM decoder network parameters

[0092] input_size output_size hidden_size num_layers 256 12000 512 1

[0093] where input_size is the input sequence size, output_size is the output sequence size, hidden_size is the hidden layer size, and num_layers is the number of network layers.

[0094] The loss function of this network is the loss function of VAE, which is composed of the reconstruction loss and the KL divergence loss, and the specific expression is:

[0095] loss = MSE(x, x') + kld_weight * KL(N(mu, sigma 2 ), N(0, 1)) (Formula 5)

[0096] Wherein, the reconstruction loss MSE(x, x') is the mean square error loss between the output sequence reconstructed by the feature extraction network and the input sequence, which measures the ability to reconstruct the input data; kld_weight is the KL divergence loss weight, KL(N(mu, sigma 2 ), N(0, 1)) is the KL divergence loss, which is used to learn the latent distribution close to the standard normal distribution, and its mathematical expression is:

[0097]

[0098] Wherein, mu is the mean of the VAE network latent distribution, and sigma is the standard deviation of the VAE network latent distribution.

[0099] Step 2: Train the classifier:

[0100] The classifier is created based on the convolutional neural network. The classifier is composed of three convolutional modules composed of two-dimensional convolution kernel, BN layer, activation function and pooling layer, and the last layer is a fully connected layer. Finally, softmax is used as the classifier output interference style classification result. The classifier network model parameters are shown in Table 3 as follows:

[0101] Table 3: Classifier network parameters

[0102]

[0103]

[0104] The first layer is the input data, and the input data size is (128, 1, 256, 1024), 128 represents 128 input sequences per batch; The channel number 1 represents a single channel, 256 is the dimension after dimension transformation, and the vector length is 1024. The subsequent convolutional modules are similar in structure and are composed of convolution layer + BN + activation function + pooling layer, but the convolution kernel size is different.

[0105] The training set data of the classifier is formed by splicing and convolution of the feature vectors of the transmitted signal extracted by the feature extraction network encoder part and the feature vectors of the echo signals interfered by different interference styles. The application is called the second training data set, that is, the tensor formed after splicing and convolution is used to train the classifier.

[0106] Let s(t) represent the transmitted signal, j(t) represent the jammed echo signal, and the feature vectors extracted by the feature extraction network encoder part are represented as y1=h_enc(s(t)) and y2=h_enc(j(t)) respectively, h_enc() represents the process of obtaining the hidden state after processing by the feature extraction network encoder, the relationship between the two feature vectors to a large extent reflects the relationship characteristics between the transmitted signal and the jammed echo signal, and are represented as:

[0107] h_enc(s(t))=Encoder(s(t)) (Formula 7)

[0108] h_enc(j(t))=Encoder(j(t)) (Formula 8)

[0109] where Encoder() represents the encoder function, which is the encoder part of the feature extraction network based on LSTM-VAE constructed by the present application.

[0110] The feature vectors y1 and y2 are spliced to obtain the spliced vector y:

[0111] y=y1+y2 (Formula 9)

[0112] The vector y is subjected to one-dimensional convolution operation for dimension transformation to obtain the final tensor Y:

[0113] Y=conv(y) (Formula 10)

[0114] where conv() represents the convolution function. Its parameter settings are shown in Table 4:

[0115] Table 4 One-dimensional convolution parameters

[0116] input_size kernel_size output_size (128,1,1024) Size = (1, 256, 3), step = 1, padding ='same' (128,256,1024)

[0117] The cross-entropy loss function is selected as the loss function in the classifier network training process. The cross-entropy loss function is generally used in classification tasks to measure the difference between two probability distributions. Let c i be the class of the radar jamming signal, p(c i ) be the target distribution, and q(c i ) be the predicted distribution. The cross-entropy of p and q can be expressed as:

[0118] H(p,q)=-∑ i p(c i )logq(c i ) (Formula 11)

[0119] Step 3: Use the trained feature extraction network and classifier to identify the jamming style.

[0120] After training, the feature extraction network model and the classifier network are obtained, and finally the overall network model is applied. First, the original transmitted signal s(t) and the received jammed echo signal j(t) are respectively encoded to obtain the encoded hidden state of the network as the feature vector of the signal. Then, the two feature vectors are spliced to obtain a spliced vector y that can represent the relationship between the transmitted signal and the jammed echo signal. Then, the spliced vector y is subjected to one-dimensional convolution operation for dimension transformation to obtain a tensor Y. Finally, the classifier is used to classify the transformed tensor to output the jamming pattern classification result.

[0121] In order to verify the performance of the method of the present application, the following tests the recognition effect of the proposed recognition network.

[0122] For the setting of the data set, in the simulation test, it is assumed that the selectable space of the transmitted waveform is composed of eight common low-interception waveforms, which are: linear frequency modulation signal (LFM), Barker code phase encoding signal, Frank code phase encoding signal, P1 code phase encoding signal, P2 code phase encoding signal, P3 code phase encoding signal, P4 code phase encoding signal, and Costas frequency modulation encoding signal. Considering the jamming pattern recognition of each frame of radar signal, the jammer jamming pattern is set to four common DRFM jamming and no jamming, a total of five: no jamming (Unjammed), smart noise convolution jamming (NCJ), intermittent sampling and repeated forwarding jamming (ISRJ), velocity deception jamming (VDJ), and range deception jamming (RDJ).

[0123] During the test training process, the training data for each transmitted signal and each jamming is 2000, and the total recognition accuracy on the training data set reaches 97%. The test data for each transmitted signal and each jamming on the test set is 400. The recognition accuracy of each transmitted signal and each jamming type recognized on the test data set when the signal-to-noise ratio is 10 dB is shown in the following table. Figure 5 According to the recognition result, it can be seen that only when the transmitted signal is P1 and the jamming signal is intermittent sampling and repeated forwarding jamming (ISRJ), the recognition rate is low, but the overall recognition accuracy is about 90%, which has good recognition effect.

[0124] Further, in order to verify the interference pattern recognition ability of the method under any change compared with the conventional method, the following comparative experiments are set: 4 kinds of transmitting signals in the data set are used for training, and the other 4 kinds of signals are tested, that is, linear frequency modulation signal (LFM), Barker code phase encoding signal, P1 code phase encoding signal and P2 code phase encoding signal are used for training, and Frank code phase encoding signal, P3 code phase encoding signal, P4 code phase encoding signal and Costas frequency modulation encoding signal are used for testing, compared with the conventional method of classifying time-frequency diagram by using CNN network under the condition of 10 dB signal-to-noise ratio, the average recognition rate curve on each interference signal is as shown in Figure 6

[0125] As can be seen from Figure 6 Compared with the conventional method, the recognition effect of the method for intermittent sampling and repeated forwarding interference (ISRJ) and RDJ is poor, but the effect for no interference (Unjammed), smart noise convolution interference (NCJ) and speed deception interference (VDJ) is better than that of the conventional method, and the overall average recognition rate is 6.5% higher than that of the conventional method, which embodies the effectiveness and robustness of the method.

[0126] The application provides a radar jamming recognition method based on LSTM and variational autoencoder, aiming at the problem that the existing jamming recognition network cannot adapt to the recognition of interference patterns under any change, a method which is not limited to the change of transmitting signals and can recognize the interference type when the transmitting signal changes is proposed.

[0127] Those skilled in the art should understand that the embodiments of the application can be provided as methods, devices (systems), computer devices or computer program products. Therefore, the application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0128] The application is described with reference to a flowchart of the method according to the embodiments of the application. It should be understood that each flow in the flowchart and the combination of the flows in the flowchart can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flow Figure 1 Each flow or multiple flows.

[0129] ​Finally, it should be noted that the above is only the preferred embodiment of the present application, and is not intended to limit the present application, although the embodiments are described in detail with reference to the present application, for those skilled in the art, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced, but any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A radar jamming recognition method based on LSTM and variational autoencoder, characterized in that, The method comprises the following steps: Step 1: training a feature extraction network An LSTM network and a VAE network are used to create a feature extraction network, which comprises an encoder, a fully connected layer and a decoder; The encoder uses an LSTM network to process the input sequence, extracts the hidden state of the feature extraction network as a feature vector, and the input sequence comprises a transmitted signal s(t) and an echo signal j(t) affected by interference, and the extracted feature vector is a transmitted feature vector and an interference feature vector; The fully connected layer comprises a linear layer, and the latent variable is constructed according to the parameter distribution of the feature extraction network hidden state; The decoder uses an LSTM network to decode, and decodes the latent variable into a reconstructed output sequence; Step 2: training a classifier A convolutional neural network is used to create a classifier, which comprises one or more convolutional modules composed of a two-dimensional convolution kernel, a BN layer, an activation function and a pooling layer, and the bottom layer of the classifier is a fully connected layer; The softmax function is used as the classifier output interference pattern classification result; Step 3: input the transmitted signal s(t) and the echo signal j(t) affected by interference to be identified, and use the trained feature extraction network and classifier to identify the interference pattern: S1: encode the transmitted signal s(t) and the received echo signal j(t) affected by interference respectively, obtain the hidden state of the encoded feature extraction network as the feature vector of the signal, and obtain the feature vector as the transmitted feature vector and the interference feature vector; S2: splice the two feature vectors in step S1 to obtain a spliced vector y that can represent the relationship between the transmitted signal s(t) and the echo signal j(t) affected by interference; S3: perform dimension transformation on the spliced vector y through one-dimensional convolution operation to obtain a tensor Y; S4: use the classifier to classify the transformed tensor Y, and output the interference pattern classification result. 2.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 1, characterized in that, The input data size of the convolutional module is (128, 1, 256, 1024), 128 represents 128 input sequences per batch; the channel number 1 represents a single channel, the dimension after dimension transformation is 256, and the vector length is 1024. 3.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 1, characterized in that, The loss function in the classifier is selected as a cross-entropy loss function, assuming that c i is a category of the radar jamming signal, p(c i ) is a target distribution, and q(c i ) is a predicted distribution. The cross-entropy of p and q is expressed as 4.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 1, characterized in that, The parameters of the feature extraction network hidden state in step 1 are the mean and log variance of the transmitted feature vector and the interference feature vector. 5.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 4, characterized in that, The training method of the feature extraction network in step 1 is centralized training: S5: combine the transmitted signal s(t) and the echo signal j(t) affected by interference of different patterns to construct an input sequence; S6: the input sequence is input into the LSTM encoder to obtain the hidden state, and the mean and log variance are extracted from the hidden state, and the latent variable is generated by reparameterization from the mean and log variance; S7: the latent variable is decoded into a reconstructed output sequence by the LSTM decoder; S8: use the mean of all signal losses between the reconstructed output sequence and the input sequence as the total loss function to train the entire feature extraction network. 6.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 4, characterized in that, The training method of the feature extraction network in step 1 is separate training: S9: combine the transmit signal s(t) and each pattern of the interfered echo signal j(t) as a separate input sequence respectively; S10: the input sequence is encoded by the long short-term memory network LSTM encoder to obtain the hidden state, the mean and the logarithmic variance are extracted from the hidden state, and the latent variable is generated by the mean and the logarithmic variance through the reparameterization; S11: the latent variable is decoded into a reconstructed output sequence by the long short-term memory network LSTM decoder; S12: the signal loss mean between the reconstructed output sequence and the input sequence in each combination is used as the corresponding loss function to train the feature extraction network of the corresponding signal.

7. The radar jamming recognition method based on LSTM and variational autoencoder according to claim 4, characterized in that, The training process of the feature extraction network includes: S13: set the input sequence as x(t), and obtain the hidden state h_enc of the feature extraction network through the long short-term memory network LSTM encoder, h_enc=Encoder(x(t)), wherein Encoder() represents the encoder function; S14: Use μ, logσ by linear layer 2 = ExtractMeanAndLogVar(h - enc) Extract the mean μ and log-variance logσ of the hidden state of the feature extraction network 2 ; S15: from the mean μ and the log variance logσ 2 The latent variable z is generated by reparameterization, z = μ + noise * σ, where noise is a random noise sampled from the standard normal distribution, and σ represents the standard deviation. S16: decode the latent variable into a reconstructed output sequence x'(t), x'(t)=Decoder(z), Decoder() represents the decoder function. 8.The radar jamming recognition method based on LSTM and variational autoencoder according to claim 4, characterized in that, The tensor Y is Y=conv(y), and conv() represents the convolution function.

9. The radar jamming recognition method based on LSTM and variational autoencoder according to claim 5 or 6, characterized in that, The loss function loss of the feature extraction network includes a mean square error loss MSE(x, x') between the output sequence reconstructed by the feature extraction network and the input sequence and a KL divergence loss, and the KL divergence loss is The loss function is loss = MSE(x, x') + kld_weight * KL(N(mu, sigma 2 ), N(0, 1)).

Images