CN117890871A - Known radar signal pre-sorting method based on long-term and short-term memory neural network - Google Patents
Known radar signal pre-sorting method based on long-term and short-term memory neural network Download PDFInfo
- Publication number
- CN117890871A CN117890871A CN202311806993.7A CN202311806993A CN117890871A CN 117890871 A CN117890871 A CN 117890871A CN 202311806993 A CN202311806993 A CN 202311806993A CN 117890871 A CN117890871 A CN 117890871A
- Authority
- CN
- China
- Prior art keywords
- vector
- neural network
- input
- pulse
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 24
- 238000000034 method Methods 0.000 title claims abstract description 17
- 230000007787 long-term memory Effects 0.000 title claims abstract description 8
- 230000006403 short-term memory Effects 0.000 title description 4
- 239000013598 vector Substances 0.000 claims abstract description 78
- 239000011159 matrix material Substances 0.000 claims abstract description 22
- 238000012545 processing Methods 0.000 claims abstract description 18
- 238000010187 selection method Methods 0.000 claims abstract description 12
- 238000013507 mapping Methods 0.000 claims abstract description 10
- 238000013075 data extraction Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 17
- 230000006835 compression Effects 0.000 claims description 12
- 238000007906 compression Methods 0.000 claims description 12
- 230000015654 memory Effects 0.000 claims description 9
- 238000013139 quantization Methods 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 5
- 238000012549 training Methods 0.000 claims description 5
- 238000001208 nuclear magnetic resonance pulse sequence Methods 0.000 abstract 1
- 238000005457 optimization Methods 0.000 description 6
- 238000011835 investigation Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010790 dilution Methods 0.000 description 1
- 239000012895 dilution Substances 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention provides a known radar signal pre-selection method based on a long-term memory neural network, which comprises the following implementation steps: firstly, processing input characteristic parameters, and splicing obtained dense vectors to obtain a joint characteristic input vector by using a single-heat coding and embedded matrix mapping coding method; then, extracting the time sequence and numerical variation association existing in the processed pulse sequence flow vector by using a GRU network-based data extraction method; classification is then achieved according to the input state vector via the fully connected network. Compared with other methods, the method solves the problem that the traditional data template matching method is difficult to cope with modulation parameters such as agility, and the algorithm is robust and can cope with pulse missing conditions.
Description
Technical Field
The invention belongs to the technical field of radar signal processing, and particularly relates to a known radar signal pre-selection method based on a long-term and short-term memory neural network.
Background
In a complex electromagnetic environment, the radar signal pulse stream received by the electronic investigation receiver is very dense, so that the radar signal to be sorted needs to be subjected to preliminary screening and dilution, i.e. pre-sorting. After pre-sorting, the pulse flow becomes sparse from dense, and then enters the main sorting treatment process, so that a sorting result can be obtained. In the pre-sorting, the sorting of known radar signals is to match the pulse streams to be sorted with radar database information in characteristic parameters, so that separation and identification are realized. The traditional pre-selection of known radar signals mainly adopts a template matching method. However, various working modes and signal parameters of the radar are complex and changeable nowadays, pulse streams received and measured by an electronic investigation receiver are dense and data are huge, the receiver is limited to self precision reasons and external environment reasons, a certain lack of leakage can exist for receiving radar pulses, and meanwhile, the problem of receiving pulse lack can be aggravated by technologies applied by some new system radars. The traditional template matching method cannot cope with changeable radar signal parameters, and can only passively cope with parameter change forms such as agility, dispersion and the like by increasing tolerance. But increasing the tolerance may allow many non-target radar signals to pass, causing sorting errors. And the pulse repetition interval verification and retrieval means in the template matching method are difficult to cope with the serious pulse missing problem.
Disclosure of Invention
The invention aims to solve the defects in the prior art, overcome the problems of complex change of parameters and pulse missing sorting in the known radar signals, and provide a pre-sorting method of the known radar signals based on a long-short-period memory neural network by considering time sequence correlation of the parameter change of the pulse;
in order to achieve the above purpose, the present invention adopts the following technical scheme: a known radar signal pre-selection method based on long-term memory neural network, comprising:
Step 1: based on the single-heat coding and the embedded matrix mapping, processing the input characteristic parameters to obtain a joint characteristic input vector;
Step 2: based on the GRU network, carrying out data extraction and training on the combined characteristic input vector to obtain a state vector h t;
Step 3: and processing the state vector h t through a fully connected network to obtain a pre-selection result.
Further, the step 1 specifically includes:
Step 1.1: setting a quantization processing range of the characteristic parameters, removing dimension units of the characteristic parameters compared with the quantization processing range, shrinking and unifying numerical ranges with extremely large phase difference among different characteristic parameters to obtain dimensionless characteristic parameters, and then carrying out digital rounding on the dimensionless characteristic parameters to realize preliminary integer-level coding to obtain a combined characteristic vector, wherein the characteristic parameters comprise carrier frequencies, pulse widths and pulse repetition intervals;
step 1.2: performing single-heat coding on the combined feature vector, and adopting embedded matrix mapping to compress sparse single-heat coding into dense coding feature vector;
step 1.3: and splicing the densely coded feature vectors to obtain a joint feature input vector.
Further, the formula for performing digital rounding on the dimensionless characteristic parameters is as follows:
where RF refers to carrier frequency, PW refers to pulse width, and PRI refers to pulse repetition interval.
Further, the formula of the densely encoded feature vector is:
epdwi=Epdwi·pdwi
Wherein pdwi is the reference to carrier frequency, pulse width, PRI characteristics, E pdwi is the mapping matrix of sparse coding feature vectors, and E pdwi is the more dense coding feature vectors after compression processing.
Further, the formula of the joint feature input vector is:
x=[epri,epw,erf]。
Where e pri is the densely encoded eigenvector of the carrier frequency after compression, e pw is the densely encoded eigenvector of the pulse width after compression, and e rf is the densely encoded eigenvector of the pulse repetition interval after compression.
Further, the step 2 specifically includes:
Step 2.1: using a GRU network of one layer as a hidden layer of the neural network;
Step 2.2: and inputting the combined characteristic input vectors into a GRU network in the neural network, and extracting, analyzing and training the relation between the combined characteristic input vectors of the pulse flow front time and the pulse flow back time through the GRU network to obtain a state vector h t.
Further, the model of the GRU network is:
Wherein x t is input information at time t; h t-1 is the output information of the hidden layer at the moment t-1; w r is a weight parameter matrix of the reset gate; w z is a weight parameter matrix of the update gate; w is a weight parameter matrix of the hidden layer output information; r t denotes reset gate control information at time t; z t is the time t update gate control information; h_r t is a reset vector indicating that reset control information resets the hidden layer output of t-1; h t' is the value of the input tanh function after the reset information is spliced with the input vector, and represents the output information of the candidate hidden layer; h t denotes updated hidden layer output information at time t; σ is sigmod functions.
Further, the step 3 specifically includes:
Step 3.1: the GRU output state vector h t is used as a fully connected network input, probability vectors of all types of known radars are output, radar signals are primarily sorted according to the probability vectors of all types of known radars, and a pre-sorting result is obtained, wherein the formula of the probability vectors of all types of known radars is as follows:
p=softmax(Woht+bo)
p=[p1,p2,…,pN]
Wherein W o is a weight matrix, b o is an offset weight vector, N represents the type of total known radar signals, and for any radar pulse input in a coded form, the probability of the radar pulse belonging to various radars is output by the full-connection network, and the sum of the probabilities in p is 1;
Step 3.2: an improved loss function based on the cross entropy principle is designed according to probability vectors of known radar models so as to approximate the network predicted value to the true value:
Where y i is the true probability distribution and p i is the hypothesized probability distribution, i.e., the value in the probability vector found previously;
step 3.3: the fully connected network is continuously trained based on the loss function minimization target in the step 3.2, and an Adam algorithm is adopted to optimize the loss function.
The beneficial effects are that: the method solves the problem that the traditional data template-based matching method is difficult to cope with modulation parameters such as agility, can cope with pulse missing conditions, considers time sequence correlation of pulse parameter variation, and has robustness.
Drawings
FIG. 1 is a flowchart of an algorithm of the present invention.
Fig. 2 is a radar signal sorting model of the present invention.
FIG. 3 is a graph comparing the optimized paths of SGD and Adam.
Detailed Description
The invention is further explained below with reference to the drawings.
As shown in fig. 1, the present invention provides a known radar signal pre-selection method based on a long-term memory neural network, which includes:
Step 1: and processing the input characteristic parameters based on the single thermal coding and the embedded matrix mapping to obtain a joint characteristic input vector.
Step 2: based on the GRU network, data extraction and training are carried out on the combined characteristic input vector, and a state vector h t is obtained.
Step 3: and processing the state vector h t through a fully connected network to obtain a pre-selection result.
In step 1, firstly, setting a quantization processing range of characteristic parameters, then removing dimension units from the characteristic parameters by comparing the quantization processing range with the quantization processing range, reducing and unifying numerical ranges with extremely large phase difference among different characteristic parameters to obtain nondimensional characteristic parameters, and then carrying out digital rounding on the nondimensional characteristic parameters to realize preliminary integer-level coding to obtain a combined characteristic vector, wherein the formula for carrying out digital rounding on the nondimensional characteristic parameters is as follows:
where RF refers to carrier frequency, PW refers to pulse width, and PRI refers to pulse repetition interval.
And then, performing one-hot coding on the combined feature vectors, and adopting embedded matrix mapping to compress sparse one-hot coding into dense coding feature vectors, wherein the formula of the dense coding feature vectors is as follows:
epdwi=Epdwi·pdwi
Wherein pdwi is the reference to carrier frequency, pulse width, PRI characteristics, E pdwi is the mapping matrix of sparse coding feature vectors, and E pdwi is the more dense coding feature vectors after compression processing.
And finally, splicing the dense coding feature vectors to obtain a joint feature input vector, wherein the formula of the joint feature input vector is as follows:
x=[epri,epw,erf]。
where e pri is the dense encoded eigenvector of the carrier frequency after compression, e pw is the dense encoded eigenvector of the pulse width after compression, and e rf is the dense encoded eigenvector of the pulse repetition interval after compression.
In this embodiment, the characteristic parameters include carrier frequency, pulse width, and pulse repetition interval.
As shown in fig. 2, in step 2, a layer of GRU network is used as a hidden layer of the neural network, then the joint feature input vector obtained in step 1 is input into the GRU network in the neural network, and the relation between the parameters before and after the pulse stream is extracted, analyzed and trained in the neural network through the GRU network to obtain a state vector h t, wherein the model of the GRU network is as follows:
Wherein x t is input information at time t; h t-1 is the output information of the hidden layer at the moment t-1; w r is a weight parameter matrix of the reset gate; w z is a weight parameter matrix of the update gate; w is a weight parameter matrix of the hidden layer output information; r t denotes reset gate control information at time t; z t is the time t update gate control information; h_r t is a reset vector indicating that reset control information resets the hidden layer output of t-1; h t' is the value of the input tanh function after the reset information is spliced with the input vector, and represents the output information of the candidate hidden layer; h t denotes updated hidden layer output information at time t; σ is sigmod functions.
In step 3, the state vector h t obtained in step 2 is input into a fully connected network, the fully connected network outputs and outputs probability vectors of classification of each model of the known radar, and radar signals are primarily classified according to the probability vectors of classification of each model of the known radar, so as to obtain a pre-classification result, wherein the formula of the probability vectors of the classification of each model of the known radar is as follows:
p=softmax(Woht+bo)
p=[p1,p2,…,pN]
Wherein W o is a weight matrix, b o is an offset weight vector, N represents the type of total known radar signals, and for any radar pulse input in a coded form, the probability of the radar pulse belonging to various radars is output by the full-connection network, and the sum of the probabilities in p is 1.
Next, an improved loss function based on the cross entropy principle is designed according to the probability vector of each model class of known radar, so as to approximate the network predicted value to the true value:
Where y i is the true probability distribution and p i is the hypothesized probability distribution, i.e., the value in the probability vector found previously.
Finally, the full-connection network is continuously trained based on a loss function minimization target, and an Adam algorithm is adopted to optimize the loss function, so that the internal parameter optimization speed and robustness of the full-connection network are improved, and the whole known radar signal pre-selection method based on the long-term and short-term memory neural network is more robust.
As shown in fig. 3, the fixed learning rate maximum gradient orientation optimization of the SGD algorithm produces a polyline-like optimization path, with slower optimization speed in the non-uniform function, and less robust iterations with fewer samples. Therefore, the optimization of the signal sorting network loss function adopts an Adam algorithm, the gradient of the model is used as a random variable, the updating direction and the learning step length in the network learning process are controlled through the high-order moment, and compared with SGD, the optimization route is smoother.
The known radar signal pre-selection method based on the long-short-term memory neural network can well cope with the problem of modulation parameters such as agility and the like under the known radar signal pre-selection scene, cope with the pulse missing condition, consider the time sequence correlation of the pulse parameter variation, and have robustness.
The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.
Claims (8)
1. A known radar signal pre-selection method based on a long-short term memory neural network, comprising:
Step 1: based on the single-heat coding and the embedded matrix mapping, processing the input characteristic parameters to obtain a joint characteristic input vector;
Step 2: based on the GRU network, carrying out data extraction and training on the combined characteristic input vector to obtain a state vector h t;
Step 3: and processing the state vector h t through a fully connected network to obtain a pre-selection result.
2. The pre-selection method of known radar signals based on long-short term memory neural network according to claim 1, wherein the step 1 specifically comprises:
Step 1.1: setting a quantization processing range of the characteristic parameters, removing dimension units of the characteristic parameters compared with the quantization processing range, shrinking and unifying numerical ranges with extremely large phase difference among different characteristic parameters to obtain dimensionless characteristic parameters, and then carrying out digital rounding on the dimensionless characteristic parameters to realize preliminary integer-level coding to obtain a combined characteristic vector, wherein the characteristic parameters comprise carrier frequencies, pulse widths and pulse repetition intervals;
step 1.2: performing single-heat coding on the combined feature vector, and adopting embedded matrix mapping to compress sparse single-heat coding into dense coding feature vector;
step 1.3: and splicing the densely coded feature vectors to obtain a joint feature input vector.
3. The pre-sorting method of known radar signals based on long-short term memory neural network according to claim 2, wherein the formula for numerical rounding of the dimensionless characteristic parameters is:
where RF refers to carrier frequency, PW refers to pulse width, and PRI refers to pulse repetition interval.
4. The method for pre-selecting known radar signals based on long-short term memory neural network according to claim 2, wherein the formula of the densely coded feature vector is:
epdwi=Epdwi·pdwi
Wherein pdwi is the reference to carrier frequency, pulse width, PRI characteristics, E pdwi is the mapping matrix of sparse coding feature vectors, and E pdwi is the more dense coding feature vectors after compression processing.
5. The known radar signal pre-selection method based on long-short term memory neural network according to claim 2, wherein the formula of the joint feature input vector is:
x=[epri,epw,erf]。
Where e pri is the densely encoded eigenvector of the carrier frequency after compression, e pw is the densely encoded eigenvector of the pulse width after compression, and e rf is the densely encoded eigenvector of the pulse repetition interval after compression.
6. The pre-selection method of known radar signals based on long-short term memory neural network according to claim 1, wherein the step2 specifically comprises:
Step 2.1: using a GRU network of one layer as a hidden layer of the neural network;
Step 2.2: and inputting the combined characteristic input vectors into a GRU network in the neural network, and extracting, analyzing and training the relation between the combined characteristic input vectors of the pulse flow front time and the pulse flow back time through the GRU network to obtain a state vector h t.
7. The known radar signal pre-selection method based on long-term memory neural network of claim 6, wherein the GRU network model is:
Wherein x t is input information at time t; h t-1 is the output information of the hidden layer at the moment t-1; w r is a weight parameter matrix of the reset gate; w z is a weight parameter matrix of the update gate; w is a weight parameter matrix of the hidden layer output information; r t denotes reset gate control information at time t; z t is the time t update gate control information; h_r t is a reset vector indicating that reset control information resets the hidden layer output of t-1; h t' is the value of the input tanh function after the reset information is spliced with the input vector, and represents the output information of the candidate hidden layer; h t denotes updated hidden layer output information at time t; σ is sigmod functions.
8. The pre-selection method of known radar signals based on long-short term memory neural network according to claim 1, wherein the step3 specifically comprises:
Step 3.1: the GRU output state vector h t is used as a fully connected network input, probability vectors of all types of known radars are output, radar signals are primarily sorted according to the probability vectors of all types of known radars, and a pre-sorting result is obtained, wherein the formula of the probability vectors of all types of known radars is as follows:
p=softmax(Woht+bo)
p=[p1,p2,…,pN]
Wherein W o is a weight matrix, b o is an offset weight vector, N represents the type of total known radar signals, and for any radar pulse input in a coded form, the probability of the radar pulse belonging to various radars is output by the full-connection network, and the sum of the probabilities in p is 1;
Step 3.2: an improved loss function based on the cross entropy principle is designed according to probability vectors of known radar models so as to approximate the network predicted value to the true value:
Where y i is the true probability distribution and p i is the hypothesized probability distribution, i.e., the value in the probability vector found previously;
step 3.3: the fully connected network is continuously trained based on the loss function minimization target in the step 3.2, and an Adam algorithm is adopted to optimize the loss function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311806993.7A CN117890871A (en) | 2023-12-26 | 2023-12-26 | Known radar signal pre-sorting method based on long-term and short-term memory neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311806993.7A CN117890871A (en) | 2023-12-26 | 2023-12-26 | Known radar signal pre-sorting method based on long-term and short-term memory neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN117890871A true CN117890871A (en) | 2024-04-16 |
Family
ID=90649963
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311806993.7A Pending CN117890871A (en) | 2023-12-26 | 2023-12-26 | Known radar signal pre-sorting method based on long-term and short-term memory neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117890871A (en) |
-
2023
- 2023-12-26 CN CN202311806993.7A patent/CN117890871A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110443818B (en) | Graffiti-based weak supervision semantic segmentation method and system | |
CN108510741B (en) | Conv1D-LSTM neural network structure-based traffic flow prediction method | |
CN109214452B (en) | HRRP target identification method based on attention depth bidirectional cyclic neural network | |
CN105913087B (en) | Object identification method based on optimal pond convolutional neural networks | |
CN108566257B (en) | Signal recovery method based on back propagation neural network | |
CN110133599B (en) | Intelligent radar radiation source signal classification method based on long-time and short-time memory model | |
CN110166484A (en) | A kind of industrial control system intrusion detection method based on LSTM-Attention network | |
CN107682319A (en) | A kind of method of data flow anomaly detection and multiple-authentication based on enhanced angle Outlier factor | |
CN111368920A (en) | Quantum twin neural network-based binary classification method and face recognition method thereof | |
CN112684427B (en) | Radar target recognition method based on serial secondary reinforcement training | |
CN109581339B (en) | Sonar identification method based on automatic adjustment self-coding network of brainstorming storm | |
CN108537790B (en) | Different-source image change detection method based on coupling translation network | |
CN115809624B (en) | Automatic analysis design method for integrated circuit microstrip line transmission line | |
CN112766603A (en) | Traffic flow prediction method, system, computer device and storage medium | |
CN111353379A (en) | Signal measurement feature matching and labeling method based on weight clustering | |
CN112765894A (en) | K-LSTM-based aluminum electrolysis cell state prediction method | |
CN111144462B (en) | Unknown individual identification method and device for radar signals | |
CN111462173A (en) | Visual tracking method based on twin network discriminant feature learning | |
CN117890871A (en) | Known radar signal pre-sorting method based on long-term and short-term memory neural network | |
CN113065395A (en) | Radar target new class detection method based on generation countermeasure network | |
CN115510915A (en) | Known radar signal sorting method based on gating cycle network principle | |
CN116189008A (en) | Remote sensing image change detection method based on fixed point number quantification | |
CN113516170B (en) | Image classification method based on Bayesian neural network random addition saturated structure | |
CN114912482A (en) | Method and device for identifying radiation source | |
CN112434716B (en) | Underwater target data amplification method and system based on condition countermeasure neural network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |