CN113156376B - SACNN-based radar radiation source signal identification method - Google Patents
SACNN-based radar radiation source signal identification method Download PDFInfo
- Publication number
- CN113156376B CN113156376B CN202110407606.7A CN202110407606A CN113156376B CN 113156376 B CN113156376 B CN 113156376B CN 202110407606 A CN202110407606 A CN 202110407606A CN 113156376 B CN113156376 B CN 113156376B
- Authority
- CN
- China
- Prior art keywords
- layer
- radiation source
- radar radiation
- representing
- data
- 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.)
- Active
Links
- 230000005855 radiation Effects 0.000 title claims abstract description 58
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 45
- 238000000605 extraction Methods 0.000 claims abstract description 33
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 43
- 230000004913 activation Effects 0.000 claims description 25
- 238000010606 normalization Methods 0.000 claims description 22
- 238000012795 verification Methods 0.000 claims description 19
- 238000004422 calculation algorithm Methods 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 17
- 238000011176 pooling Methods 0.000 claims description 15
- 238000013178 mathematical model Methods 0.000 claims description 12
- 230000004044 response Effects 0.000 claims description 9
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 230000007704 transition Effects 0.000 claims description 4
- 230000015556 catabolic process Effects 0.000 claims description 3
- 238000006731 degradation reaction Methods 0.000 claims description 3
- 238000001514 detection method Methods 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000003213 activating effect Effects 0.000 claims description 2
- 238000006243 chemical reaction Methods 0.000 abstract description 3
- 238000013527 convolutional neural network Methods 0.000 description 20
- 230000000694 effects Effects 0.000 description 6
- 238000004088 simulation Methods 0.000 description 5
- 238000010200 validation analysis Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 238000012567 pattern recognition method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/021—Auxiliary means for detecting or identifying radar signals or the like, e.g. radar jamming signals
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a SACNN-based radar radiation source signal identification method, which comprises the following steps: 1) Constructing a data set; 2) Preprocessing data; 3) Constructing an SRNN local feature extraction module; 4) Constructing an Attention module; 5) Constructing a CNN global feature extraction module; 6) Training a radar radiation source identification network; 7) And identifying radar radiation source signals. The network structure constructed by the invention directly adopts the one-dimensional time domain radar radiation source signals for training, can directly extract the characteristics of the one-dimensional time domain radar radiation source signals and identify the characteristics, and solves the problems that the radar radiation source signal identification accuracy is low under the condition of low signal to noise ratio in the existing radar radiation source identification method, the existing two-dimensional time-frequency image-based radar radiation source signal identification method needs time-frequency conversion, the time consumption is more, and the instantaneity is poor.
Description
Technical Field
The invention relates to the technical field of electronic countermeasure, in particular to a SACNN-based radar radiation source signal identification method.
Background
Radar radiation source identification (Radar Emitter Identification, REI) is to analyze and process intercepted enemy radar signals to obtain working parameters and signal characteristic parameters of the enemy radar, and compare the working parameters with a known radar database to judge the type, working mode and position of the radar so as to master information such as a fight platform, working state and threat level and provide information support for battlefield electromagnetic situation sensing, threat warning, fight planning and the like. With the increasing complexity of the electromagnetic environment of the battlefield, the conventional recognition method based on pulse descriptors (Pulse Description Words, PDW) parameters cannot well meet the requirements of radar radiation source signal recognition under the condition of low signal-to-noise ratio. While the advent of low probability of interception (low probability of intercept, LPI) radar makes radar radiation source signal identification more difficult. Therefore, the method has very important practical significance for accurately identifying the radar radiation source signals.
The key to radar radiation source identification is feature extraction. In recent years, radar radiation source identification technology based on machine learning is widely focused by researchers due to its stronger generalization and intelligence. As an important research branch in the machine learning field, deep learning and application thereof are research hotspots in the artificial intelligence field, and have achieved good effects in fields such as machine translation, question answering, image classification, speech recognition, text classification, and the like. Deep learning differs most from traditional pattern recognition methods in that it is capable of automatically extracting features from data. By means of layer-by-layer feature transformation, features of the sample in the original space are transformed into a new feature space, and complex classification tasks can be completed by using a simple model, so that classification or prediction is easier.
Many scholars at home and abroad introduce a deep learning method into radar radiation source identification so as to achieve better identification effect than the traditional manual identification method. Wan J et al propose a recognition method based on CNN-TPOT to recognize the two-dimensional time-frequency diagram, and the overall recognition rate of 12 signals reaches 94.42% under the condition that the signal-to-noise ratio is-4 dB; zhang M et al propose a hybrid classifier comprising two relatively independent sub-networks of Convolutional Neural Network (CNN) and Elman Neural Network (ENN), with a signal-to-noise ratio of-2 dB, the overall recognition rate of 12 signals reaching 94.5%; guo Q et al propose to utilize the recognition method of deep convolutional network transfer learning, turn signal into time-frequency chart and precondition, input to CNN pretraining model carry on the characteristic extraction, get the classification result with SVM classifier finally, under the condition that the signal-to-noise ratio is-2 dB, the overall recognition rate can reach 97% to 9 kinds of modulating signals.
The main problems of the method are that: first, under low snr conditions, the recognition accuracy is not high, and many of the results mentioned in the above documents are obtained under higher snr conditions, which are difficult to achieve in a battlefield electromagnetic environment; second, the recognition accuracy of various signals is unbalanced, and the signals with unobvious characteristics and difficult recognition are the signals most likely to be adopted by enemies and have the greatest threat, which may have serious consequences, thereby limiting the practical application of the networks.
Disclosure of Invention
The invention provides a SACNN-based radar radiation source signal identification method, which aims at the problems that the radar radiation source signal identification accuracy is low under the condition of low signal-to-noise ratio in the prior art, the existing two-dimensional time-frequency image-based radar radiation source signal identification method needs time-frequency conversion, the time consumption is more, and the instantaneity is poor.
A SACNN-based radar radiation source signal identification method comprises the following steps:
1) Constructing a dataset
Sampling radar radiation source signals detected by the detection equipment, intercepting a fixed length, taking the fixed length as data, and marking a label;
building a training set, a verification set and a test set: the data and the labels are randomly disturbed correspondingly, and a training set, a verification set and a test set are divided according to the proportion;
2) Data preprocessing
Carrying out normalization processing on data of each generated training set, verification set and test set by using a Min-Max normalization algorithm;
the mathematical model of the Min-Max normalization algorithm is expressed as follows:
wherein Input representing Min-Max normalization algorithm, < +.> and />Respectively the minimum and maximum of all input data
Normalizing the response of the algorithm for Min-Max;
carrying out one-hot coding on the labels of all data;
slicing the normalized data twice, wherein the data with the shape (1024, 1) is sliced into the data with the shape (8,128,1) at the first time, and the data with the shape (8,128,1) is sliced into the data with the shape (8,8,16,1) at the second time;
3) And (3) constructing an SRNN local feature extraction module:
constructing a 3-layer SRNN local feature extraction module, and inputting data with the shape of (8,8,16,1);
the module network structure is as follows:
the input of the 0 th layer is the minimum subsequence with the length of 16, each subsequence corresponds to the standard RNN structure, and the output is the hidden state of the 0 th layer;
wherein ,represents +.0 on layer>A personal hidden state; />Indicate->A circulation unit of the layer; />Represents the smallest subsequence of layer 0; />The value is 16 for the minimum subsequence length of the 0 th layer; the output of layer 0 is；
The input of layer 1 is the hidden state of the output of layer 0From hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 1, i.e. +.>；
wherein ,represents +.1 on layer>A personal hidden state; />A circulation unit representing a first layer; />The value is 16 for the minimum subsequence length of the 0 th layer;
the input of layer 2 is the hidden state of the output of layer 1From hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 2, i.e. +.>That is, the output of the 3-layer SRNN local feature extraction module;
wherein ,representing the output of a 3-layer SRNN local feature extraction module; />Represents +.2 on layer>A personal hidden state;a circulation unit representing layer 2; />The value of the minimum subsequence length of the layer 1 is 128;
in the 3-layer SRNN local feature extraction module constructed above, the circulating units are all gated circulating units GRU; in the GRU, the output unit is set to 32, and the activation function is set to tanh;
the mathematical model of the activation function tanh is expressed as follows:
4) Constructing an Attention module;
5) Constructing a CNN global feature extraction module;
6) Training a radar radiation source identification network;
7) Identifying radar radiation source signals;
the step 4) of constructing an Attention module comprises the following steps:
an Attention module is built, the structure of the Attention module is totally 4 layers, and the Attention module is sequentially: a first Permute layer, a full connection layer, a second Permute layer, a multiplexing layer, wherein the first and second Permute layers Permute the input dimensions according to a given pattern; the activation function of the fully-connected layer is softmax; the multiplexing layer multiplies the output of the second Permute layer by the input of the first Permute layer in bits, the result being the output of the Attention module.
The implementation of the attention state transition is represented as follows:
wherein F represents the output of the Attention module;representing the normalized matching degree; />Representing a hidden state of the input; />Representing an attention scoring mechanism; /> and />A weight coefficient representing the moment i; />The offset corresponding to the moment i;
the 5) constructing a CNN global feature extraction module, which comprises the following steps:
a11-layer CNN global feature extraction module is built, and the structure of the CNN global feature extraction module is as follows: first batch of normalization layers, first convolution layers, first maximum pooling layers, second batch of normalization layers, second convolution layers, second maximum pooling layers, third batch of normalization layers, third convolution layers, third maximum pooling layers, flattening layers and full connection layers;
setting the number of convolution kernels in the first to third convolution layers as 32, setting the convolution kernel size as 4 multiplied by 1, setting the step length as 1, and setting the activation function as ReLU; the sizes of the cores of the pooling areas of the first to third maximum pooling layers are 2 multiplied by 1, and the step sizes are all set to be 1; the number of neurons of the full-connection layer is set to 8, and the activation function is softmax;
the mathematical model of the activation function ReLU is expressed as follows:
wherein Representing the response of the input value x of the network after passing the activation function ReLU.
The mathematical model of the activation function softmax is expressed as follows:
wherein Indicate->Value of individual element->Indicating all->Sum of values of individual elements->Response to the activation function softmax;
step 6) training a radar radiation source identification network, comprising:
inputting the preprocessed training set sample into a training network in a radar radiation source identification network, and verifying the training result of each round by using the preprocessed verification set sample;
updating the network weight by adopting an Adam algorithm;
the Adam algorithm is as follows:
wherein Expressed as a loss function->Is a gradient of (2); />Expressed as iteration weights; />Representing a gradient operator; />Represent +.0 initialized>Is a first order moment estimate of (2); />Denoted as initialized to 0->Is determined by the second moment estimate of (2); />Exponential decay for first moment estimationThe ratio is 0.9; />The exponential decay rate estimated for the second moment is 0.999; />Representing a transpose operation;for learning rate, the initial setting is 0.001 in the present invention; />Is a smooth constant, the divisor is prevented from being 0, and the value is +.>；
Adopting a cross entropy loss function; to avoid the occurrence of overfitting to prevent degradation of the generalization ability of the network;
the cross entropy loss function is expressed as follows:
wherein H (p, q) represents a cross entropy loss function; p (x) i ) Representing the true distribution of the samples; q (x) i ) Representing the distribution predicted by the model; the smaller the cross entropy loss function, the closer the true distribution of the sample is to the distribution predicted by the model;
introducing early stop braking, taking the loss of the verification set as a standard, and stopping training when the loss of the verification set is not reduced for 10 continuous rounds;
introducing learning rate attenuation, and setting the minimum learning rate to be 0;
setting the maximum training round number as 200 rounds and the batch_size as 200;
taking the identification accuracy of the verification set as a standard, and storing a network model with highest identification accuracy by calling a function ModelCheckPoint;
the data set of the modulation type signal is: two-phase encoded signals, chirped continuous wave signals, costas signals, frank signals and multiphase codes P1, P2, P3, P4;
each modulation type signal generates 2000 sample signals under the conditions of signal-to-noise ratio of { -20dB, -18dB, -16dB, -14dB, -12dB, -10dB, -8dB, -6dB, -4dB, -2dB,0dB,2dB,4dB,6dB,8dB and 10dB, namely each modulation signal generates 32000 samples in total, eight different modulation type signals generate 256000 samples in total, and the sampling point number of each sample is 1024.
Compared with the prior art, the invention has the following advantages:
firstly, the SRNN structure is introduced into the radar radiation source identification field and improved, and the defects of low training speed and poor instantaneity of the conventional radar radiation source identification technology based on the standard RNN structure are overcome. The SRNN multi-layer structure can effectively extract the signal characteristics of the local radar radiation source in a short time and the signal characteristics of the global radar radiation source in a long time, and the characteristic extraction is more sufficient. On the basis, compared with the SRNN, the SRNN+attention+CNN method provided by the invention has the advantages that the recognition accuracy and the training speed are further improved, the recognition accuracy is improved by 1%, and the total training time is only 74.5% of that of the SRNN.
The invention provides a SACNN-based radar radiation source signal identification method, which comprises the following steps: 1) Constructing a data set; 2) Preprocessing data; 3) Constructing an SRNN local feature extraction module; 4) Constructing an Attention module; 5) Constructing a CNN global feature extraction module; 6) Training a radar radiation source identification network; 7) And identifying radar radiation source signals. The network structure constructed by the invention directly adopts the one-dimensional time domain radar radiation source signals for training, can directly extract the characteristics of the one-dimensional time domain radar radiation source signals and identify the characteristics, and solves the problems that the radar radiation source signal identification accuracy is low under the condition of low signal to noise ratio in the existing radar radiation source identification method, the existing two-dimensional time-frequency image-based radar radiation source signal identification method needs time-frequency conversion, the time consumption is more, and the instantaneity is poor.
Drawings
FIG. 1 is a SRNN+Attention+CNN model structure;
FIG. 2 is a SRNN network architecture;
FIG. 3 is an Attention layer structure;
FIG. 4 shows the recognition accuracy of the SRNN+Attention+CNN model on 8 signals under different signal-to-noise ratios.
Detailed Description
Example 1
A radar radiation source signal identification method based on SRNN+attention+CNN comprises the following steps:
1) Constructing a dataset
Sampling radar radiation source signals detected by the detection equipment, intercepting a fixed length, taking the fixed length as data, and marking a label;
and generating a data set comprising two-phase coded signals, linear frequency modulation continuous wave signals, costas signals, frank signals and multiphase codes P1-P4, wherein the data set comprises eight different modulation types of signals. The signal parameters are shown in table 1 below:
each modulation type signal generates 2000 sample signals under the conditions of signal-to-noise ratio of { -20dB, -18dB, -16dB, -14dB, -12dB, -10dB, -8dB, -6dB, -4dB, -2dB,0dB,2dB,4dB,6dB,8dB and 10dB, namely each modulation signal generates 32000 samples in total, and eight different modulation type signals generate 256000 samples in total. The number of samples per sample is 1024.
A training set, a validation set, and a test set are established. The data and label correspondence were randomly scrambled, and training, validation and test sets were partitioned in proportion (60%, 20%).
2) Data preprocessing
Preprocessing the generated training set, verification set and test set data.
And carrying out normalization processing on the data of each generated training set, verification set and test set by using a Min-Max normalization algorithm.
The mathematical model of the Min-Max normalization algorithm is expressed as follows:
wherein Input representing Min-Max normalization algorithm, < +.> and />Respectively the minimum and maximum of all input data
The response of the algorithm is normalized for Min-Max.
All data tags were one-hot encoded. The one-hot encoding converts attribute values into binary features for discrete features, with only one bit being valid at any time.
And carrying out slicing treatment on the normalized data twice. The data with the shape (1024, 1) is split into the data with the shape (8,128,1) for the first time, and the data with the shape (8,128,1) is split into the data with the shape (8,8,16,1) for the second time, so that the requirement of the SRNN network on the dimension of input data is met.
3) And (3) constructing an SRNN local feature extraction module:
a3-layer SRNN local feature extraction module is built, and the network structure is shown in figure 2. The input is data in the shape (8,8,16,1) after two slicing operations.
The input of the 0 th layer is the minimum subsequence with the length of 16, each subsequence corresponds to the standard RNN structure, and the output is the hidden state of the 0 th layer;
wherein ,represents +.0 on layer>A personal hidden state; />Indicate->A circulation unit of the layer; />Represents the smallest subsequence of layer 0; />The value is 16 for the minimum subsequence length of the 0 th layer; the output of layer 0 is；
The input of layer 1 is the hidden state of the output of layer 0From hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 1, i.e. +.>；
wherein ,represents +.1 on layer>A personal hidden state; />A circulation unit representing a first layer; />The value is 16 for the minimum subsequence length of the 0 th layer;
the input of layer 2 is the hidden state of the output of layer 1From hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 2, i.e. +.>That is, the output of the 3-layer SRNN local feature extraction module;
wherein ,representing the output of a 3-layer SRNN local feature extraction module; />Represents +.2 on layer>A personal hidden state;a circulation unit representing layer 2; />The value of the minimum subsequence length of the layer 1 is 128;
in the 3-layer SRNN local feature extraction module constructed above, the circulating units are all gate-controlled circulating units (GRUs). In the GRU, the output unit is set to 32 and the activation function is set to tanh.
The mathematical model of the activation function tanh is expressed as follows:
4) Constructing an Attention module:
the attention mechanism (Attention mechanism) is essentially a weight probability distribution mechanism that focuses more on finding useful information in the input data that is significantly related to the current output, exploring auto-correlation in the signal, highlighting part of the features related to the prediction, and thus improving the quality of the output, making training more efficient. The network structure is shown in fig. 3.
An Attention module is built, the structure of the Attention module is totally 4 layers, and the Attention module is sequentially: a first Permute layer, a full connection layer, a second Permute layer, and a multiple layer. Wherein the first and second Permute layers Permute the dimensions of the input according to a given pattern; the activation function of the fully-connected layer is softmax; the multiplexing layer multiplies the output of the second Permute layer by the input of the first Permute layer in bits, the result being the output of the Attention module.
The implementation of the attention state transition is represented as follows:
wherein F represents the output of the Attention module;representing the normalized matching degree; />Representing a hidden state of the input; />Representing an attention scoring mechanism; /> and />A weight coefficient representing the moment i; />The corresponding offset at time i. Through the above three formulas, the transition of the attentiveness state can be realized.
5) Constructing a CNN global feature extraction module:
a11-layer CNN global feature extraction module is built, and the structure of the CNN global feature extraction module is as follows: first batch of normalization layers, first convolution layers, first maximum pooling layers, second batch of normalization layers, second convolution layers, second maximum pooling layers, third batch of normalization layers, third convolution layers, third maximum pooling layers, flattening layers and full connection layers.
Setting the number of convolution kernels in the first to third convolution layers as 32, setting the convolution kernel size as 4 multiplied by 1, setting the step length as 1, and setting the activation function as ReLU; the sizes of the cores of the pooling areas of the first to third maximum pooling layers are 2 multiplied by 1, and the step sizes are all set to be 1; the number of neurons in the fully connected layer was set to 8 and the activation function was softmax.
The mathematical model of the activation function ReLU is expressed as follows:
wherein Representing the response of the input value x of the network after passing the activation function ReLU.
The mathematical model of the activation function softmax is expressed as follows:
wherein Indicate->Value of individual element->Indicating all->Sum of values of individual elements->For activating the response of the function softmax.
6) Training radar radiation source identification network:
and inputting the preprocessed training set sample into a training network in the radar radiation source identification network, and verifying the training result of each round by using the preprocessed verification set sample.
And updating the network weight by adopting an Adam algorithm.
The Adam algorithm is as follows:
wherein Expressed as a loss function->Is a gradient of (2); />Expressed as iteration weights; />Representing a gradient operator; />Represent +.0 initialized>Is a first order moment estimate of (2); />Denoted as initialized to 0->Is determined by the second moment estimate of (2); />The exponential decay rate estimated for the first moment is 0.9; />The exponential decay rate estimated for the second moment is 0.999; />Representing a transpose operation;for learning rate, the initial setting is 0.001 in the present invention; />To smooth outConstant, divisor 0, value +.>。
Adopting a cross entropy loss function; to avoid the occurrence of overfitting to prevent degradation of the generalization ability of the network.
The cross entropy loss function is expressed as follows:
wherein H (p, q) represents a cross entropy loss function; p (x) i ) Representing the true distribution of the samples; q (x) i ) Representing the distribution predicted by the model. The smaller the cross entropy loss function, the closer the true distribution of the sample is to the distribution predicted by the model.
An early shutdown system was introduced, taking the loss of the validation set as a standard, and training was stopped when the loss of the validation set was not reduced for 10 consecutive rounds.
And introducing learning rate attenuation, and when the loss of the verification set is not reduced by 3 continuous rounds, reducing the learning rate to 10% of the original learning rate, and setting the minimum learning rate to 0.
The maximum training round number is set to 200 rounds and the batch_size is 200.
And taking the identification accuracy of the verification set as a standard, and storing a network model with highest identification accuracy by calling a function ModelCheckPoint.
7) Identification of radar radiation source signals
And calling a stored network model, and sequentially inputting the preprocessed test set data into a trained radar radiation source identification network to obtain an identification result of each sample of the test set.
Example 2
The effects of the present invention are further described below in conjunction with simulation experiments:
1. simulation conditions
The hardware platform and the software platform adopted by the simulation experiment of the invention are shown in the following table 2.
2. Simulation experiment and result analysis
The radar radiation source identification simulation experiment provided by the invention is to identify the modulation type of each radar radiation source signal by adopting the SRNN+attention+CNN method provided by the invention, and count the total number of samples accurately identified by eight modulation type signals under each signal-to-noise ratio to obtain the accurate identification rate. The recognition accuracy of the SRNN+attention+CNN model on 8 signals under different signal-to-noise ratios is shown in figure 4.
It can be seen that the accuracy rate of 100% can be basically achieved under the condition that the signal-to-noise ratio is greater than or equal to-10 dB; the identification accuracy of all signals is more than 60% at-20 dB, the BPSK and FMCW accuracy is more than 70%, and the accuracy of other 5 multiphase code signals is more than 85%, which proves that the SRNN+attention+CNN method provided by the invention has better identification effect on radar radiation source signals under the condition of low signal-to-noise ratio.
Based on the above experiments, the srnn+attention+cnn model was further compared with SRNN and some classical models. The training effect of the network is measured by comparing the loss of the test set with the identification accuracy, the number of training rounds, the total training time and the test set test time. The experimental results are shown in table 3 below. The experimental results were averaged over multiple experiments.
As can be seen from the experimental results in Table 3, compared with the SRNN, the SRNN+attention+CNN method provided by the invention has the advantages that the recognition accuracy and the training speed are further improved, the recognition accuracy is improved by 1%, and the total training time is only 74.5% of that of the SRNN. Compared with GRU of standard RNN structure, the method has obvious advantages in training speed and greater advantages in test set loss and recognition accuracy. Compared with other classical models, the method has certain advantages in recognition accuracy, is only worse than AlexNet in training time, is leveled with VGG16, and is better than VGG19 and ResNet 18. In terms of test time, the SRNN+Attention+CNN model is close to AlexNet, and has great advantages compared with other classical models. In summary, the method has better recognition effect and good noise immunity.
Claims (5)
1. A SACNN-based radar radiation source signal identification method comprises the following steps:
1) Constructing a dataset
Sampling radar radiation source signals detected by the detection equipment, intercepting a fixed length, taking the fixed length as data, and marking a label;
building a training set, a verification set and a test set: the data and the labels are randomly disturbed correspondingly, and a training set, a verification set and a test set are divided according to the proportion;
2) Data preprocessing
Carrying out normalization processing on data of each generated training set, verification set and test set by using a Min-Max normalization algorithm;
the mathematical model of the Min-Max normalization algorithm is expressed as follows:
where x represents the input of the Min-Max normalization algorithm, min (x) and Max (x) are the minimum and maximum values of all input data respectively,
x' is the response of the Min-Max normalization algorithm;
carrying out one-hot coding on the labels of all data;
slicing the normalized data twice, wherein the data with the shape (1024, 1) is sliced into the data with the shape (8,128,1) at the first time, and the data with the shape (8,128,1) is sliced into the data with the shape (8,8,16,1) at the second time;
3) And (3) constructing an SRNN local feature extraction module:
constructing a 3-layer SRNN local feature extraction module, and inputting data with the shape of (8,8,16,1);
the module network structure is as follows:
the input of the 0 th layer is the minimum subsequence with the length of 16, each subsequence corresponds to the standard RNN structure, and the output is the hidden state of the 0 th layer;
wherein ,representing the t hidden state on layer 0; />A circulation unit representing layer 0; mss the smallest subsequence of layer 0; l (L) 0 The value is 16 for the minimum subsequence length of the 0 th layer; the output of layer 0 is->The input of layer 1 is the hidden state of layer 0 output +.>From the hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 1, i.e. +.>
wherein ,representing the t hidden state on layer 1; />A circulation unit representing a first layer; l (L) 0 The value is 16 for the minimum subsequence length of the 0 th layer;
the input of layer 2 is the hidden state of the output of layer 1From the hidden state->The composed subsequence corresponds to standard RNN structure and is output as hidden state of layer 2, i.e. +.>Namely the output of the 3-layer SRNN local feature extraction module;
wherein F represents the output of the 3-layer SRNN local feature extraction module;representing the t hidden state on layer 2; />A circulation unit representing layer 2; l (L) 1 The value of the minimum subsequence length of the layer 1 is 128;
in the 3-layer SRNN local feature extraction module constructed above, the circulating units are all gated circulating units GRU; in the GRU, the output unit is set to 32, and the activation function is set to tanh;
the mathematical model of the activation function tanh is expressed as follows:
4) Constructing an Attention module;
5) Constructing a CNN global feature extraction module;
6) Training a radar radiation source identification network;
7) And identifying radar radiation source signals.
2. The SACNN-based radar radiation source signal identification method according to claim 1, characterized in that: the step 4) of constructing an Attention module comprises the following steps: an Attention module is built, the structure of the Attention module is totally 4 layers, and the Attention module is sequentially: a first Permute layer, a full connection layer, a second Permute layer, a multiple layer; wherein the first and second Permute layers Permute the dimensions of the input according to a given pattern; the activation function of the fully-connected layer is softmax; the multiplexing layer multiplies the output of the second Permute layer by the input of the first Permute layer in a para-position way, and the result is used as the output of the Attention module;
the implementation of the attention state transition is represented as follows:
wherein F represents the output of the Attention module; a, a i Representing the normalized matching degree; h is a i Representing a hidden state of the input; e, e i Representing an attention scoring mechanism; and Wi A weight coefficient representing the moment i; b i The corresponding offset at time i.
3. The SACNN-based radar radiation source signal identification method according to claim 2, characterized in that:
the 5) constructing a CNN global feature extraction module, which comprises the following steps:
a11-layer CNN global feature extraction module is built, and the structure of the CNN global feature extraction module is as follows: first batch of normalization layers, first convolution layers, first maximum pooling layers, second batch of normalization layers, second convolution layers, second maximum pooling layers, third batch of normalization layers, third convolution layers, third maximum pooling layers, flattening layers and full connection layers;
setting the number of convolution kernels in the first to third convolution layers as 32, setting the convolution kernel size as 4 multiplied by 1, setting the step length as 1, and setting the activation function as ReLU; the sizes of the cores of the pooling areas of the first to third maximum pooling layers are 2 multiplied by 1, and the step sizes are all set to be 1; the number of neurons of the full-connection layer is set to 8, and the activation function is softmax;
the mathematical model of the activation function ReLU is expressed as follows:
where f (x) represents the response of the input value x of the network after passing the activation function ReLU;
the mathematical model of the activation function softmax is expressed as follows:
wherein ei Represents the value of the ith element, Σ j e j Representing the sum of the values of all j elements, S i For activating the response of the function softmax.
4. A SACNN-based radar radiation source signal identification method according to claim 1, 2 or 3, characterized in that:
step 6) training a radar radiation source identification network, comprising:
inputting the preprocessed training set sample into a training network in a radar radiation source identification network, and verifying the training result of each round by using the preprocessed verification set sample;
updating the network weight by adopting an Adam algorithm;
the Adam algorithm is as follows:
m←β 1 m+(1-β 1 )g
v←β 2 v+(1-β 2 )g 2
wherein g is represented as a gradient of a loss function L (θ); θ is represented as an iteration weight;representing a gradient operator; m represents a first moment estimate of g initialized to 0; v is represented as a second moment estimate of g initialized to 0; beta 1 The exponential decay rate estimated for the first moment is 0.9; beta 2 The exponential decay rate estimated for the second moment is 0.999; t represents a transpose operation; alpha is learning rate, and is initially set to 0.001; epsilon is a smooth constant, the divisor is prevented from being 0, and the value is 10 -9 ；
Adopting a cross entropy loss function; to avoid the occurrence of overfitting to prevent degradation of the generalization ability of the network;
the cross entropy loss function is expressed as follows:
wherein H (p, q) represents a cross entropy loss function; p (x) i ) Representing the true distribution of the samples; q (x) i ) Representing the distribution predicted by the model; the smaller the cross entropy loss function, the representationThe closer the true distribution of the sample is to the distribution predicted by the model;
introducing early stop braking, taking the loss of the verification set as a standard, and stopping training when the loss of the verification set is not reduced for 10 continuous rounds;
introducing learning rate attenuation, and setting the minimum learning rate to be 0;
setting the maximum training round number as 200 rounds and the batch_size as 200;
and taking the identification accuracy of the verification set as a standard, and storing a network model with highest identification accuracy by calling a function ModelCheckPoint.
5. The SACNN-based radar radiation source signal identification method as claimed in claim 4, wherein: the data set of the modulation type signal is: two-phase encoded signals, chirped continuous wave signals, costas signals, frank signals and multiphase codes P1, P2, P3, P4;
each modulation type signal generates 2000 sample signals under the conditions of signal-to-noise ratio of { -20dB, -18dB, -16dB, -14dB, -12dB, -10dB, -8dB, -6dB, -4dB, -2dB,0dB,2dB,4dB,6dB,8dB and 10dB, namely each modulation signal generates 32000 samples in total, eight different modulation type signals generate 256000 samples in total, and the sampling point number of each sample is 1024.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110407606.7A CN113156376B (en) | 2021-04-15 | 2021-04-15 | SACNN-based radar radiation source signal identification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110407606.7A CN113156376B (en) | 2021-04-15 | 2021-04-15 | SACNN-based radar radiation source signal identification method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113156376A CN113156376A (en) | 2021-07-23 |
| CN113156376B true CN113156376B (en) | 2023-08-11 |
Family
ID=76868044
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110407606.7A Active CN113156376B (en) | 2021-04-15 | 2021-04-15 | SACNN-based radar radiation source signal identification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113156376B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114021459A (en) * | 2021-11-05 | 2022-02-08 | 西安晟昕科技发展有限公司 | Identification method of small sample radar radiation source |
| CN114048774B (en) * | 2021-11-10 | 2024-08-02 | 厦门大学 | Resnet communication radiation source identification method and system based on se-block |
| CN114358050B (en) * | 2021-12-11 | 2024-02-02 | 中国人民解放军国防科技大学 | Intelligent radar radiation source identification method based on bicubic interpolation and WVD feature square matrix |
| CN114726692B (en) * | 2022-04-27 | 2023-06-30 | 西安电子科技大学 | SERESESESENet-LSTM-based radiation source modulation mode identification method |
| CN114814776B (en) * | 2022-06-24 | 2022-10-14 | 中国空气动力研究与发展中心计算空气动力研究所 | PD radar target detection method based on graph attention network and transfer learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109271926A (en) * | 2018-09-14 | 2019-01-25 | 西安电子科技大学 | Intelligent Radiation source discrimination based on GRU depth convolutional network |
| WO2020248471A1 (en) * | 2019-06-14 | 2020-12-17 | 华南理工大学 | Aggregation cross-entropy loss function-based sequence recognition method |
| CN112098957A (en) * | 2020-09-15 | 2020-12-18 | 西安电子科技大学 | Complex radar radiation source identification method based on one-dimensional self-walking convolution neural network |
| CN112115924A (en) * | 2020-09-27 | 2020-12-22 | 西安电子科技大学 | Radar radiation source identification method based on one-dimensional CNN and LSTM |
| JP6830707B1 (en) * | 2020-01-23 | 2021-02-17 | 同▲済▼大学 | Person re-identification method that combines random batch mask and multi-scale expression learning |
-
2021
- 2021-04-15 CN CN202110407606.7A patent/CN113156376B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109271926A (en) * | 2018-09-14 | 2019-01-25 | 西安电子科技大学 | Intelligent Radiation source discrimination based on GRU depth convolutional network |
| WO2020248471A1 (en) * | 2019-06-14 | 2020-12-17 | 华南理工大学 | Aggregation cross-entropy loss function-based sequence recognition method |
| JP6830707B1 (en) * | 2020-01-23 | 2021-02-17 | 同▲済▼大学 | Person re-identification method that combines random batch mask and multi-scale expression learning |
| CN112098957A (en) * | 2020-09-15 | 2020-12-18 | 西安电子科技大学 | Complex radar radiation source identification method based on one-dimensional self-walking convolution neural network |
| CN112115924A (en) * | 2020-09-27 | 2020-12-22 | 西安电子科技大学 | Radar radiation source identification method based on one-dimensional CNN and LSTM |
Non-Patent Citations (1)
| Title |
|---|
| 一种深度学习的雷达辐射源识别方法;李昆;朱卫纲;;电子设计工程(第12期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113156376A (en) | 2021-07-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113156376B (en) | SACNN-based radar radiation source signal identification method | |
| Wu et al. | Object detection based on RGC mask R‐CNN | |
| Frizzi et al. | Convolutional neural network for smoke and fire semantic segmentation | |
| CN110109060B (en) | Radar radiation source signal sorting method and system based on deep learning network | |
| CN109495214B (en) | Channel coding type identification method based on one-dimensional inclusion structure | |
| CN110929842B (en) | Accurate intelligent detection method for non-cooperative radio signal burst time region | |
| CN114881092A (en) | Signal modulation identification method based on feature fusion | |
| CN115238749B (en) | Modulation recognition method based on feature fusion of transducer | |
| CN114636975A (en) | LPI radar signal identification method based on spectrogram fusion and attention mechanism | |
| Zhang et al. | Open set recognition of communication signal modulation based on deep learning | |
| CN111553186A (en) | Electromagnetic signal identification method based on depth long-time and short-time memory network | |
| Zhou et al. | Fast minutiae extractor using neural network | |
| Ding et al. | Efficient BiSRU combined with feature dimensionality reduction for abnormal traffic detection | |
| CN116797796A (en) | Signal identification method based on time-frequency analysis and deep learning under DRFM intermittent sampling | |
| Zhang et al. | Turning wingbeat sounds into spectrum images for acoustic insect classification | |
| Zhang et al. | Advjnd: Generating adversarial examples with just noticeable difference | |
| Xia et al. | Combination of multi‐scale and residual learning in deep CNN for image denoising | |
| CN114296041A (en) | Radar radiation source identification method based on DCNN and Transformer | |
| CN116471154A (en) | Modulation signal identification method based on multi-domain mixed attention | |
| CN117233706B (en) | Radar active interference identification method based on multilayer channel attention mechanism | |
| Hui et al. | FoolChecker: A platform to evaluate the robustness of images against adversarial attacks | |
| CN117131436A (en) | Radiation source individual identification method oriented to open environment | |
| CN116243248A (en) | Multi-component interference signal identification method based on multi-label classification network | |
| Gao et al. | Supervised Contrastive Learning‐Based Modulation Classification of Underwater Acoustic Communication | |
| Gai et al. | Spectrum sensing method based on residual cellular 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |