PRI dithering signal pulse sequence searching method based on Mogrimer LSTM

Abstract

The invention discloses a PRI jitter signal pulse sequence searching method based on Mogrifier LSTM, and the method comprises the steps: generating radar radiation source signal pulse data, each pulse data sample comprises RF, PW, PA, TOA and DOA, and forming a time sequence radar signal pulse sample set S; a PRI value is calculated for each radar radiation source signal pulse sample in the time sequence radar signal pulse set S, normalization processing is carried out on RF, PW, TOA, PA, DOA and PRI of each sample to form a sample feature x, and a label value of each sample is a corresponding radar pulse signal group number; a radar signal pulse sequence search model based on the MogrifierLSTM network is constructed by taking the sample feature x as input and the radar pulse signal group number as output; and inputting the pulse data sample into the trained radar signal pulse sequence search model to obtain a radar radiation source pulse sequence prediction result. The method is higher in accuracy and good in stability, and can be used for radar radiation source pulse sequence search in a complex electromagnetic environment.

Description

Technical Field

[0001] The present invention relates to radar signal sorting technology, in particular to a method for searching pulse sequences of PRI jitter signals based on Mogrifier-LSTM. Background Art

[0002] Radar signal sorting and recognition is the core content of modern warfare electronic warfare. Adopting advanced recognition technology can identify the modulation mode and parameters of the enemy radar first, so as to obtain the purpose of the target radar and judge its threat level.

[0003] With the rapid development of technology, radars with various complex modulation modes have emerged one after another. The current electromagnetic environment and space environment have become increasingly harsh, which poses a huge challenge to radar signal sorting and recognition. How to accurately complete radar signal sorting in real time in the increasingly harsh electromagnetic environment is the key point and difficulty in the research field of radar reconnaissance.

[0004] The direct sequence retrieval method model of traditional signal sorting is simple, with high probabilities of batch addition and batch omission, and it is difficult to realize the retrieval of radar pulses with complex systems. At present, many scholars are also researching how to extract more detailed features of radar pulse signals through neural networks. However, in the field of radar signal sorting, the currently commonly used algorithms are still some clustering algorithms of machine learning and the application of self-organizing networks, and the application of deep learning algorithms is less. Summary of the Invention

[0005] The purpose of the present invention is to propose a method for searching pulse sequences of jitter PRI signals based on an improved long short-term memory network to solve the problem of searching for jitter PRI signal sequences that are difficult to handle by traditional algorithms.

[0006] The technical solution for realizing the purpose of the present invention is as follows: A method for searching pulse sequences of PRI jitter signals based on Mogrifier LSTM, the steps are as follows:

[0007] Step 1, generate radar emitter signal pulse data, each pulse data sample includes RF, PW, PA, TOA and DOA, and form a time-series radar signal pulse sample set S;

[0008] Step 2, calculate the PRI value for each radar emitter signal pulse sample in the time-series radar signal pulse set S, perform normalization processing on RF, PW, TOA, PA, DOA, and PRI of each sample, form sample features x, and the label value of each sample is the corresponding radar pulse signal group number;

[0009] Step 3, extract a training sample set S 1 and a validation sample set S 2 ;

[0010] Step 4: Using the sample feature x as the input and the radar pulse signal group number as the output, construct a radar signal pulse sequence search model based on the Mogrifier LSTM network. By adding the pulse sample feature x at the current moment t and the output h of the Mogrifier LSTM model at the previous moment t-1 for interaction, enhance the context relationship of the pulse sequence;

[0011] Step 5: Set the parameters of the Mogrifier LSTM network, set the batch size and learning rate, and use the training set and validation set to train the radar signal pulse sequence search model;

[0012] Step 6: Input the pulse data sample into the trained radar signal pulse sequence search model to obtain the radar emitter pulse sequence prediction result.

[0013] Furthermore, in Step 1, generate radar emitter signal pulse data. Each pulse data sample includes RF, PW, PA, TOA, and DOA, forming a time-series radar signal pulse sample set S. The specific method is as follows:

[0014] Generate 100 groups of jitter signals with jitters of 5%, 10%, 20%, and 30% respectively. The PRI value of each group of signals is a random value within [100, 400] s, and the length of each group of data is 10,000.

[0015] Furthermore, in Step 4, using the sample feature x as the input and the radar pulse signal group number as the output, construct a radar signal pulse sequence search model based on the Mogrifier LSTM network, where:

[0016] The Mogrifier-LSTM network includes an input gate, a forget gate, an output gate, a memory unit, and a prediction layer. Based on the LSTM, an interaction space for the pulse sample feature x at the current moment t and the output h of the Mogrifier LSTM model at the previous moment t-1 is added. That is, after the pulse sample feature x at the current moment t is transformed through the sigmoid function, the control state u t is obtained. Then, the control state u t and the output h of the Mogrifier LSTM model at the previous moment t-1 are dot-multiplied, so that each element in the output h of the Mogrifier LSTM model at the previous moment t-1 is transformed to different degrees:

[0017] u t = σ(W u·x(t) + b u )

[0018] h′ t-1 = 2h t-1 *u t

[0019] where W u , b u are the weight and bias term respectively;

[0020] The output h′ of the Mogrifier LSTM model after interaction t-1 is processed by sigmoid to obtain the control state v t , and the control state v t is used to transform each element in the pulse sample feature x t at the current moment to obtain the pulse sample feature x' after interaction t , v t and x' t The update process is as follows:

[0021] v t = σ(W v ·h t-1 + b v )

[0022] x′ t = 2x t *v t

[0023] where W v , b v are the weight and bias term respectively;

[0024] The output h′ of the Mogrifier LSTM model after interaction t-1 , the pulse sample feature x' after interaction t are used as the input of the LSTM, and the output h t at the current moment is calculated,

[0025] i t = σ(W xi x′ t + h′ t-1 W hi + b i )

[0026] f t = σ(W xf x′ t + h′ t-1 W hf + b f )

[0027] ot = σ(W xo x′ t + h′ t-1 W ho + b o )

[0028]

[0029] h t = o t ⊙ C t

[0030] where i t , f t , o t are the input gate, forget gate, and output gate information respectively, is the candidate memory cell, C t is the memory cell, W xi , W hi , W xf , W hf , W xo , W ho , W xc , W hc are the weights, b i , b f , b o , b c are the bias terms;

[0031] Finally, the prediction layer calculates the radar pulse signal group number y using the sotfmax function based on the output h t at the current time:

[0032] y = softmax(h t ).

[0033] Furthermore, in step 5, set the MogrifierLSTM network parameters, set the batch size and learning rate, and train the radar signal pulse sequence search model using the training set and validation set, where:

[0034] Set the number of neural network layers to 2; set the number of memory cell nodes in the first layer to 256, the number of memory cell nodes in the second layer to 256, and the forget gate function for each node; set the temporal prediction classification function of the prediction layer to the Softmax function; set the optimization algorithm in the Mogrifier LSTM network model to the adam optimization algorithm, the loss function to the cross-entropy function, and the activation functions to the rectified linear unit activation function and the hyperbolic tangent activation function.

[0035] Further, in step 5, set the parameters of the MogrifierLSTM network, set the batch size and the learning rate, and use the training set and the validation set to train the radar signal pulse sequence search model, where:

[0036] The batch size is set to 128; the learning rate is set to 0.0001. Input the training set and the validation set into the MogrifierLSTM network for 100 rounds of iterative training.

[0037] A PRI jitter signal pulse sequence search system based on Mogrifier LSTM implements the above-mentioned PRI jitter signal pulse sequence search method based on Mogrifier LSTM to realize the PRI jitter signal pulse sequence search based on Mogrifier LSTM, and six modules respectively execute steps 1 to 6.

[0038] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned PRI jitter signal pulse sequence search method based on Mogrifier LSTM to realize the PRI jitter signal pulse sequence search based on Mogrifier LSTM.

[0039] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, it implements the above-mentioned PRI jitter signal pulse sequence search method based on Mogrifier LSTM to realize the PRI jitter signal pulse sequence search based on Mogrifier LSTM.

[0040] Compared with the prior art, the significant advantages of the present invention are as follows: 1) Build a Mogrifier-LSTM for automatically extracting the features of time-series radar signals, which overcomes the problems of complex models and the need for feature extraction of radar emitter signals in the existing methods, realizes the end-to-end sequence retrieval of time-series radar emitter signals, overcomes the disadvantage of the need for a large amount of prior experience in feature extraction of radar emitter signals in the prior art, and reduces the computational complexity of radar emitter signal classification. 2) The Mogrifier-LSTM network used retains the integrity of the time series of radar emitter signals, overcomes the disadvantage of the traditional method that cannot fully explore the long-time features of signals, and improves the effect of radar emitter signal sequence retrieval. 3) The Mogrifier-LSTM network is adopted to extract the sequence features of radar emitter signal features. Compared with the ordinary LSTM network, two control gates are added, which enhances the mutual screening of the previous output and the current input, and better extracts the high-dimensional hidden sequence features of the emitter. Description of the Drawings

[0041] Figure 1 is the flowchart of the radar emitter pulse sequence prediction based on Mogrifier-LSTM of the present invention.

[0042] Figure 2 is the structural diagram of the LSTM network unit.

[0043] Figure 3 is the structural diagram of the Mogrifier LSTM network unit.

[0044] Figure 4 is the curve graph of the prediction accuracy rate of the model training sequence.

[0045] Figure 5 is the curve graph of the loss function value during the model training process. Specific implementation manners

[0046] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0047] A method for searching the pulse sequence of a jittered PRI signal based on Mogrifier-LSTM, the steps are as follows:

[0048] Step 1, generate radar emitter signal pulse data. Each pulse data sample includes carrier frequency (RF), pulse width (PW), amplitude (PA), time of arrival (TOA), azimuth (DOA), and form a time-series radar signal pulse sample set S;

[0049] 100 groups of jitter signals with jitters of 5%, 10%, 20%, and 30% are respectively generated. The PRI value of each group of signals is a random value within [100, 400] s, and the length of each group of data is 10,000.

[0050] Step 2, calculate the PRI value of the TOA of each radar emitter signal pulse sample in the time-series radar signal pulse set S. The label value of each sample is the corresponding radar pulse signal group number. Normalize the RF, PW, TOA, PA, DOA, and PRI of each sample to form the sample feature x, so as to prevent some data from occupying too much weight during the training of the network and affecting the prediction result, thereby obtaining the preprocessed radar signal pulse sequence sample set S'.

[0051] Step 3, extract the first 70% of the signals from the preprocessed radar signal pulse sequence sample set S' to form the training sample S 1 , extract the first 20% of the samples from the remaining 30% of the radar signals to form the verification sample set S 2 , and the remaining 10% of the radar signals are used as the test sample set S3 。

[0052] Step 4: Build a Mogrifier-LSTM network including an input gate, a forget gate, an output gate, a memory unit, and a prediction layer to extract the features of radar time-series signals, filter out invalid information, avoid gradient disappearance, and search for the radar signal pulse sequence;

[0053] Without changing the original structure of LSTM, MogrifierLSTM only adds the interaction space of the pulse sample feature x t at the current moment and the output h t-1 of the MogrifierLSTM model at the previous moment. After x t is transformed by the sigmoid function, the control state u t is obtained. Multiply u t and h t-1 to make each element in h t-1 transformed to different degrees:

[0054] u t =σ(W u ·x(t)+b u )+

[0055] h′ t-1 =2h t-1 *u t

[0056] where W u , b u are the weight and bias terms respectively.

[0057] h′ t-1 is processed by sigmoid to obtain the control state v t . Use v t to implement the transformation of each element in x t to obtain the value x' t after interaction. The update processes of v t and x' t are as follows:

[0058] v t =σ(W v ·h t-1 +b v )

[0059] x′ t =2x t *v t

[0060] where W v , b v are the weight and bias terms respectively.

[0061] Take the h′ obtained from the above interaction t-1 and x′ t as the input of the LSTM to calculate the output h at the current moment t ,

[0062] i t = σ(W xi x′ t + h′ t-1 W hi + b i )

[0063] f t = σ(W xf x′ t + h′ t-1 W hf + b f )

[0064] o t = σ(W xo x′ t + h′ t-1 W ho + b o )

[0065]

[0066] h t = o t ⊙ C t

[0067] where i t , f t , o t are the input gate, forget gate, and output gate information respectively, is the candidate memory cell, C t is the memory cell, W xi , W hi , W xf , W hf , W xo , W ho , W xc , W hc are the weights, b i , b f , b o , b c are the bias terms;

[0068] Finally, the prediction layer calculates the class number y of the pulse sequence using the softmax function based on the output h at the current moment: t y = softmax(h

[0069] ) t )

[0070] Step 5: Set the number of neural network layers to 2; set the number of memory unit nodes in the first layer to 256, the number of memory unit nodes in the second layer to 256, and the forgetting gate function for each node; set the temporal prediction classification function of the prediction layer to the Softmax function; set the optimization algorithm in the Mogrifier LSTM network model to the adam optimization algorithm, the loss function to the cross-entropy function, the activation function to the rectified linear unit activation function and the hyperbolic tangent activation function.

[0071] Set the learning rate of the Mogrifier-LSTM network, and input the training sample set S 1 and the validation sample set S 2 into this network, and perform iterative training to obtain a trained Mogrifier-LSTM network model and save the model parameters;

[0072] Step 6: Input the samples in the test sample set S 3 into the trained Mogrifier-LSTM network model to obtain the predicted class results of each sample sequence.

[0073] The present invention also proposes a PRI jitter signal pulse sequence search system based on Mogrifier LSTM, implements the PRI jitter signal pulse sequence search method based on Mogrifier LSTM, and realizes the PRI jitter signal pulse sequence search based on Mogrifier LSTM.

[0074] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the PRI jitter signal pulse sequence search method based on Mogrifier LSTM, and realizes the PRI jitter signal pulse sequence search based on Mogrifier LSTM.

[0075] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the PRI jitter signal pulse sequence search method based on Mogrifier LSTM, and realizes the PRI jitter signal pulse sequence search based on Mogrifier LSTM.

[0076] Embodiment

[0077] To verify the effectiveness of the solution of the present invention, the following experiment is carried out.

[0078] The PRI jitter signal pulse sequence search method based on Mogrifier LSTM is as follows:

[0079] Step 1: Data Generation and Preprocessing

[0080] Generate radar emitter signal pulse data, preprocess the data, and divide the dataset. 100 groups of jitter signals with jitters of 5%, 10%, 20%, and 30% are generated respectively. The PRI value of each group of signals is a random value within [100, 400] s. The length of each group of data is 10,000. The first 7,000 pulses are used as the training set S, the first 2,000 pulses of the remaining pulses are used for the validation set V, and the last 1,000 pulses are used as the test set T..

[0081] Calculate the PRI value for the TOA of each radar emitter signal pulse sample in the sequential radar signal pulse set S. The label value of each sample is the corresponding radar pulse signal group number. Normalize RF, PW, TOA, PA, DOA, and PRI for each sample to form the sample feature x.

[0082] Step 2: Construct a Pulse Sequence Prediction Model Based on MogrifierLSTM

[0083] On the basis of not changing the original structure of LSTM, MogrifierLSTM only adds the pulse sample feature x at the current moment t and the output h of the MogrifierLSTM model at the previous moment t-1 of the interaction space. x t After being transformed by the sigmoid function, the control state u is obtained t . u t and h t-1 are dot-multiplied so that each element in h t-1 is transformed to different degrees:

[0084] u t = σ(W u ·x(t)+b u )

[0085] h′ t-1 = 2h t-1 *u t

[0086] where W u and b u are the weight and bias terms respectively.

[0087] h′ t-1 After being sigmoid-processed, the control state v is obtained t , and v t is used to realize the transformation of each element in x t to obtain the interacted value x' t , v t and x' tThe update process is as follows:

[0088] v t = σ(W v ·h t-1 + b v )

[0089] x′ t = 2x t * v t

[0090] where W v and b v are the weight and bias terms respectively.

[0091] Taking h′ t-1 and x′ t obtained from the above interaction as the inputs of the LSTM to calculate the output h t ,

[0092] i t = σ(W xi x′ t + h′ t-1 W hi + b i )

[0093] f t = σ(W xf x′ t + h′ t-1 W hf + b f )

[0094] o t = σ(W xo x′ t + h′ t-1 W ho + b o )

[0095]

[0096] h t = o t ⊙ C t

[0097] where i t , f t , o t are the input gate, forget gate, and output gate information respectively, is the candidate memory cell, C t is the memory cell, W xi , W hi , W xf , W hf , Wxo , W ho , W xc , W hc are weights, and b i , b f , b o , b c are bias terms;

[0098] The final prediction layer outputs h at the current moment t , and uses the softmax function to calculate the class number y of the pulse sequence:

[0099] y = softmax(h t )

[0100] Step 3: Network parameter setting

[0101] Set the number of Mogrifier-LSTM neurons to 2

[0102] Set the number of nodes in the first-layer memory unit to 256, the number of nodes in the second-layer long short-term memory unit to 256, and the forgetting layer function of each node to softmax;

[0103] Set the prediction layer to a time series prediction classification function of Softmax;

[0104] Set the optimization algorithm in the Mogrifier-LSTM network model to the adam optimization algorithm, the loss function to the cross-entropy function, and the activation functions to the rectified linear unit activation function and the hyperbolic tangent activation function.

[0105] Step 4: Network training optimization:

[0106] Set the batch size to 128; set the learning rate to 0.0001. Input the training set and the validation set into the Mogrifier LSTM network, and after 100 rounds of iterative training, save the trained model parameters. As Figure 4 , 5 shown, at the 55th round, the model tends to converge, and the sequence prediction accuracy reaches 96.13%.

[0107] Step 5: Test and verification

[0108] Input the test sample set S 3 into the trained Mogrifier-LSTM network model to obtain the sequence prediction classification result.

[0109] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0110] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A PRI jitter signal pulse sequence search method based on Mogrifier LSTM, characterized in that: Here are the steps: Step 1, generate radar radiation source signal pulse data, each pulse data sample contains RF, PW, PA, TOA and DOA, forming a time series radar signal pulse sample set S; Step 2: Calculate the PRI value for each radar emitter signal pulse sample in the time series radar signal pulse set S, normalize the RF, PW, TOA, PA, DOA, and PRI of each sample to form a sample feature x, and the label value of each sample is the corresponding radar pulse signal group number; Step 3, extracting training samples S1 and verification sample set S2 from the preprocessed radar signal pulse sequence sample set S'; Step 4: Taking the sample feature x as input and the radar pulse signal group number as output, a radar signal pulse sequence search model based on the MogrifierLSTM network is constructed. By adding the pulse sample feature x at the current moment t And the MogrifierLSTM model output h at the previous moment t-1 interaction, enhancing the contextual relationship of the pulse sequence; Step 5, set the MogrifierLSTM network parameters, set the batch number and learning rate, and use the training set and validation set to train the radar signal pulse sequence search model; Step 6: Input the pulse data samples into the trained radar signal pulse sequence search model to obtain the radar radiation source pulse sequence prediction result.

2. The PRI jitter signal pulse sequence search method based on Mogrifier LSTM according to claim 1, characterized in that: Step 1: Generate radar radiation source signal pulse data. Each pulse data sample contains RF, PW, PA, TOA and DOA, forming a time series radar signal pulse sample set S. The specific method is as follows: 100 groups of jitter signals with jitter of 5%, 10%, 20%, and 30% are generated respectively. The PRI value of each group of signals is a random value within [100,400]s, and the length of each group of data is 10000.

3. The PRI jitter signal pulse sequence search method based on Mogrifier LSTM according to claim 1, characterized in that: Step 4: Taking the sample feature x as input and the radar pulse signal group number as output, a radar signal pulse sequence search model based on the MogrifierLSTM network is constructed, where: The Mogrifier-LSTM network includes an input gate, a forget gate, an output gate, a memory unit, and a prediction layer. On the basis of LSTM, the current pulse sample feature x is added. t And the MogrifierLSTM model output h at the previous moment t-1 The interaction space, that is, the current moment pulse sample feature x t After the sigmoid function conversion, the control state u is obtained t , and then the control state u t And the MogrifierLSTM model output h at the previous moment t-1 Perform point multiplication so that the MogrifierLSTM model outputs h at the previous moment t-1 The elements in are transformed to varying degrees: u t =σW u ·x(t)+b u ) h' t-1 =2h t-1 *u t Where W u , b u are weight and bias terms respectively; Output the interactive MogrifierLSTM model h' t-1 After sigmoid processing, the control state v is obtained t , using the control state v t Realize the pulse sample feature x at the current moment t Transform each element in to obtain the pulse sample feature x' after interaction t , v t and x' t The update process is as follows: v t =σ(W v ·h t-1 +b v ) x’ t =2x t *v t Where W v , b v are weight and bias terms respectively; Output the interactive MogrifierLSTM model h' t-1 , the pulse sample characteristics x' after interaction t As the input of LSTM, the output h at the current moment is calculated t , i t =σ(W xi x’ t +h’ t-1 W hi +b i ) f t =σ(W xf x’ t +h’ t-1 W hf +b f ) o t =σ(W xo x’ t +h’ t-1 W ho +b o ) h t =o t ⊙C t Among them, i t 、f t , o t They are the input gate, forget gate, and output gate information respectively. is a candidate memory unit, C t is the memory unit, W xi , W hi , W xf , W hf , W xo , W ho , W xc , W hc is the weight, b i , b f , b o , b c is the bias term; Finally, the prediction layer outputs h according to the current moment t , use the sotfmax function to calculate the radar pulse signal group number y: and softmax(h) t )。 4. The PRI jitter signal pulse sequence search method based on Mogrifier LSTM according to claim 1, characterized in that: Step 5: Set the MogrifierLSTM network parameters, batch size and learning rate, and use the training set and validation set to train the radar signal pulse sequence search model, where: Set the number of neural network layers to 2; set the number of memory unit nodes in the first layer to 256, the number of memory unit nodes in the second layer to 256 and the forget gate function of each node; set the prediction layer time series prediction classification function to the Softmax function; set the optimization algorithm in the Mogrifier LSTM network model to the adam optimization algorithm, the loss function to the cross entropy function, and the activation function to the linear rectified unit activation function and the hyperbolic tangent activation function.

5. The PRI jitter signal pulse sequence search method based on Mogrifier LSTM according to claim 1, characterized in that: Step 5: Set the MogrifierLSTM network parameters, batch size and learning rate, and use the training set and validation set to train the radar signal pulse sequence search model, where: The batch size is set to 128; the learning rate is set to 0.0001, and the training set and validation set are input into the MogrifierLSTM network for 100 rounds of iterative training.

6. A PRI jitter signal pulse sequence search system based on Mogrifier LSTM, characterized in that: The PRI jitter signal pulse sequence search method based on Mogrifier LSTM as described in any one of claims 1-5 is implemented to realize the PRI jitter signal pulse sequence search based on Mogrifier LSTM, and steps 1 to 6 are respectively performed in six modules.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for searching a PRI jitter signal pulse sequence based on Mogrifier LSTM according to any one of claims 1 to 5 is implemented to realize searching a PRI jitter signal pulse sequence based on Mogrifier LSTM.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for searching a PRI jitter signal pulse sequence based on Mogrifier LSTM according to any one of claims 1 to 5 is implemented to realize searching a PRI jitter signal pulse sequence based on Mogrifier LSTM.

Images