CN114201987A - Active interference identification method based on self-adaptive identification network - Google Patents
Active interference identification method based on self-adaptive identification network Download PDFInfo
- Publication number
- CN114201987A CN114201987A CN202111318774.5A CN202111318774A CN114201987A CN 114201987 A CN114201987 A CN 114201987A CN 202111318774 A CN202111318774 A CN 202111318774A CN 114201987 A CN114201987 A CN 114201987A
- Authority
- CN
- China
- Prior art keywords
- sample
- ajs
- output
- ajsarnet
- signal
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000012360 testing method Methods 0.000 claims abstract description 47
- 238000012549 training Methods 0.000 claims abstract description 44
- 238000004088 simulation Methods 0.000 claims abstract description 26
- 230000009466 transformation Effects 0.000 claims abstract description 7
- 238000011056 performance test Methods 0.000 claims abstract description 5
- 239000000523 sample Substances 0.000 claims description 163
- 239000013598 vector Substances 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 7
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000005457 optimization Methods 0.000 claims description 3
- 239000012468 concentrated sample Substances 0.000 claims description 2
- 238000013507 mapping Methods 0.000 claims description 2
- 230000003044 adaptive effect Effects 0.000 description 8
- 230000006870 function Effects 0.000 description 7
- 238000010801 machine learning Methods 0.000 description 5
- 238000005259 measurement Methods 0.000 description 4
- 238000013508 migration Methods 0.000 description 3
- 238000013526 transfer learning Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Signal Processing (AREA)
- Complex Calculations (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an active interference identification method based on a self-adaptive identification network, and belongs to the technical field of active interference identification. The method comprises the following steps: 1) constructing an active interference signal simulation data set by using time-frequency transformation; 2) dividing an active interference signal simulation data set to obtain a sample set, a query set, a support set and a test set; 3) building an AJSARNET with input as a sample pair and outputting a relation score; 4) initializing AJSARNET; 5) training the AJSARNET initialized in the step 4) based on the sample set and the query set to obtain the trained AJSARNET; 6) and testing the trained AJSARNET based on the support set and the test set to obtain a performance test result. The method improves robustness or generalization capability, and AJS recognition accuracy and efficiency.
Description
Technical Field
The invention relates to an active interference identification method based on a self-adaptive identification network, and belongs to the technical field of active interference identification.
Background
Active Jamming (AJ) identification refers to a technology of identifying the type of a jammer transmitting the signal in real time by analyzing characteristic parameters of active jamming intercepted by an Electronic Support Measurement (ESM) system, further analyzing the transmitter operating mode of the jammer and evaluating the corresponding threat level. The active interference identification technology plays an increasingly important role in the fields of radar target detection and radar anti-interference, and is widely concerned by researchers in recent years.
With the rapid development of machine learning technology, especially deep learning, compared with the traditional active interference identification methods such as the early signal parameter matching method and the expert system method, the method based on machine learning structurally improves the active interference identification method, effectively improves the identification performance of active interference signals, and gradually becomes a main technical means in the field of active interference identification.
However, in the face of the current increasingly complex electromagnetic environment, machine learning-based active interference identification also faces a number of challenges: 1) due to military or commercial confidentiality, the active interference sample is difficult to acquire, and a large amount of time and cost are consumed in the acquisition process, so that the scale of a data set for training, testing and evaluating the active interference identification model is limited, the overfitting phenomenon is easy to occur in the training process of the model, and the accuracy of active interference identification is reduced; 2) the electromagnetic environment is complex, noise from electronic equipment and an external signal source can pollute active interference, the identification performance of the active interference is seriously reduced, a machine learning model is particularly easy to influence, and a noise point input by the model can change the optimization direction of model parameters, so that the model generates wrong output, namely, the active interference identification method based on machine learning generally has the weakness of low robustness; 3) the active interference recognition model generally has a complex network structure, which includes a large number of parameters to be trained and optimized, and has high computational complexity and high time cost.
Disclosure of Invention
The invention aims to solve the problems of low accuracy and poor real-time performance of active interference identification on the premise of low interference-to-noise ratio (JNR), and provides an active interference identification method based on a self-adaptive identification network, which relies on a novel end-to-end active interference self-adaptive identification network AJSARNET based on a meta-migration learning strategy and small sample learning (few-shot learning, FSL), acquires knowledge from a few samples through generalization and simulation of human learning capacity, and realizes active interference identification.
In order to achieve the above purpose, the present invention adopts the following technical scheme.
The active interference identification method based on the self-adaptive identification network relies on an AJS self-adaptive identification network model, namely AJSARNET;
the AJSARNET is formed by cascading an embedded module and a relation module;
the input of the embedded module is the input of AJSARNET, and the embedded module learns the high-dimensional feature expression of the input sample pair and outputs an effective embedded vector;
the relation module comprises two convolution layers and two full-connection layers, and the last full-connection layer outputs a relation score to determine a final recognition result;
the input of the relation module is the output of the embedded module, namely: a valid embedding vector; the relationship module learns the mapping between the effective embedding vector output by the embedded module and the relationship score of the input sample pair;
the identification method comprises the following steps:
step 1: constructing an active interference signal simulation data set by using time-frequency transformation;
wherein, active jamming signal is abbreviated as AJS;
specifically, the active interference signal simulation data set generated in step 1 is used for training, testing and evaluating the AJS recognition model, and includes N × S × M samples, and each sample is composed of two parts: 1) a two-dimensional time-frequency spectrogram is generated after the AJS signal is subjected to time-frequency transformation; 2) the label of the AJS signal is the true value of the AJS recognition result of the signal;
wherein N refers to the type number of the AJS signals in the constructed AJS analog data set; s means that samples contained in each type of AJS signal comprise S different interference-to-noise ratio levels; m means that each class of AJS signals contains M samples at each dry-to-noise ratio level;
wherein, the dry-to-noise ratio refers to the energy ratio of signal to noise, i.e. Jamming-to-noise ratio, abbreviated as JNR, and is defined by formula (1):
wherein, PjAnd PnInterference power and noise power, respectively;
step 1.1, constructing an AJS simulation data set, which specifically comprises the following steps: according to the requirement of the AJS recognition task, simulating N multiplied by S multiplied by M AJS signals to generate an AJS signal set;
the simulation of the AJS signal is specifically determined by a modulation mode, a carrier frequency, a pulse width, a bandwidth and a sampling frequency, and comprises the following substeps:
step 1.1A, selecting N different modulation modes, and determining N types of the AJS signals;
step 1.1B, according to the N types of the AJS signals determined in the step 1.1A, setting different signal parameters for each type of the AJS signals by using N different modulations, respectively generating S multiplied by M AJS analog signals, and generating an original AJS signal set;
step 1.1C, S dry-to-noise ratio levels are set, and white noise with S amplitudes is added to the AJS signals in the original AJS signal set generated in step 1.1B according to formula (1) to generate an AJS signal set;
specifically, white noise with the same amplitude is added to every M signal samples of each type of AJS signal, and each type of AJS signal contains S uniform different dry-to-noise ratio levels;
step 1.2, performing time-frequency transformation on each AJS signal in the AJS signal set output in the step 1.1 to generate a corresponding two-dimensional time-frequency spectrogram, taking the type of the AJS signal as a label of the AJS signal, and forming an active interference signal simulation data set by using the AJS two-dimensional time-frequency spectrogram and the label of the AJS signal;
step 2: dividing the active interference signal simulation data set output in the step 1 according to the set identification problem of the C-way K-shot to obtain a sample set, a query set, a support set and a test set of the AJS identification model;
the C-way K-shot recognition problem specifically refers to a classification task constructed in meta learning; the sample set and the query set jointly form a training set for training the AJS recognition model; the supporting set and the testing set are used for testing the AJS recognition model, the supporting set contains C-class data, and each class of data consists of K samples; the class C data specifically refers to a set formed by samples of which the types and the dry-to-noise ratio levels of the AJS signals are the same in the AJS analog data set output in the step 1;
the method specifically comprises the following substeps of dividing the active interference signal analog data set output in the step 1:
step 2.1, randomly selecting a C-type AJS signal in the active interference signal simulation data set output in the step 1, and dividing the C-type AJS signal to form a support set and a test set;
the specific dividing method comprises the following steps: for each type of the C type AJS signals, randomly selecting K samples to generate a support set, and generating a test set by the rest samples;
step 2.2, removing samples belonging to the support set or the test set generated in the step 2.1 from the active interference signal simulation data set output in the step 1, and forming a training set by the remaining samples;
step 2.3, dividing the training set output in the step 2.2 to form a sample set and a query set;
the specific dividing method comprises the following steps: randomly selecting K samples from each class of the training set output in the step 2.2 to generate a sample set, and generating a query set by using the rest samples in the training set;
and step 3: according to the requirements of the AJS recognition task, building AJSARNET;
the input of the AJSARNET comprises a sample pair formed by two AJS two-dimensional time-frequency spectrograms; the output of the AJSARNET is the relation score of the input sample pair;
wherein, the relationship score is a number from 0 to 1, and is used for measuring the similarity of two samples in the sample pair: the larger the relation score is, the more similar the two samples are, and when the relation score is larger than a certain threshold value, the two samples can be judged to be the same type of AJS signals;
and 4, step 4: initializing the AJSARNET established in the step 3;
initializing AJSARNET, specifically assigning an initial value to each parameter of a neural network;
and 5: training the initialized AJSARNET output in the step 4 based on the sample set and the query set output in the step 2 to obtain the trained AJSARNET;
specifically, setting L element training iteration units epicode for the initialized AJSARNET training output in the step 4;
for the set C-way K-shot recognition task, setting an epsilon to send all samples of the selected meta-training set into a network to be trained, and completing a forward calculation and back propagation process for each sample, specifically: selecting C classes in a training set formed by the sample set and the query set output in the step 2, training a primary model by using all samples of the selected C class data, and updating parameters of the primary model; selecting other C class training models for the next epicode; for the sample set containing NxS-C class data and the query set output in step 2, each epoch contains
An epicode;
when the NxS-C is not evenly divisible by C, the last epadiode of the current epoch represents that the type of the untrained data is less than the type C, and at the moment, the C type needs to be randomly selected and supplemented from the trained data to train the last epadiode;
specifically, the training process of each epicode includes the following sub-steps:
step 5.1, respectively taking out C-type AJS signals with the same type from the sample set and the query set output in the step 2 to generate a sample pair;
the sample set output in the step 2 and the query set respectively contain K samples and M-K samples for each type of AJS signals;
in particular, by S respectivelyijAnd QijDenotes the j-th sample, S, of the i-th signal of the class C AJS signal taken from the sample set and the query setijAnd QijThe expression of (c) is specifically shown in formula (2) and formula (3);
Sij＝(xij，li)，i＝1，2，…，C，j＝1，2，…，K (2)
Qij＝(yij，li)，i＝1，2，…，C，j＝1，2，…，M-K (3)
wherein x and y respectively represent a two-dimensional time-frequency spectrogram of a sample set and a sample of a query set, and l represents a label of the sample;
the generated sample pair is specifically shown as formula (4);
Pijmn＝(Sij，Qmn) (4)
wherein, i is 1, 2, …, C, j is 1, 2, …, K; m-1, 2, …, C, n-1, 2, …, M-K; that is, each sample in the sample set taken in step 6.1 forms a sample pair with each sample in the query set taken in step 6.1, and each epamode generates C × K × C × (M-K) sample pairs;
step 5.2, sequentially inputting the sample pairs output in the step 5.2 into AJSARNET to be trained according to the types of the sample sets of the sample pairs for calculation to obtain relationship scores;
wherein, AJSARNET to be trained specifically means: if the epsilon is the first epsilon of the model training stage, the initialized AJSARNET output in the step 4 is obtained, otherwise, the initialized AJSARNET output by the last epsilon is obtained;
step 5.2, specifically comprising the following substeps:
step 5.2.1 groups the sample pairs output in step 5.2 to obtain a sample pair group Timn；
Wherein, i is 1, 2, …, C, M is 1, 2, …, C, n is 1, 2, …, M-K, i.e. the generated sample pair group has C × (M-K) × C;
grouping the sample pairs specifically includes: sample a query set by QmnAnd samples S of a certain type of sample setijK sample pairs of 1, 2, …, K are used as a sample pair group TimnAs shown in formula (5);
Timn＝{Pijmn|j＝1，2，…，K}＝{(Sij，Qmn)|j＝1，2，…，K} (5)
step 5.2.2 pairs the samples output in step 5.2.1 against group TimnSequentially inputting the embedded modules of AJSARNET to be trained for calculation to obtain the total embedded vector of each sample pair group;
wherein, i is 1, 2, …, C, M is 1, 2, …, C, n is 1, 2, …, M-K;
specifically, the embedded module of the AJSARNET to be trained can pair the input samples into a group TimnOutputs a valid embedded vector, i.e. a sample pair group TimnGenerating K effective embedding vectors, wherein the sum of the K effective embedding vectors is TimnThe total embedded vector of (1);
step 5.2.3, sequentially inputting the total embedded vector output in the step 5.2.2 into a relation module of AJSARNET to be trained to obtain a relation score of each sample pair group;
specifically, the relationship score of each sample pair group describes the similarity between a certain query set sample and a certain sample set sample;
step 5.3, substituting the relation score output in the step 5.2 into a loss function formula to calculate the training loss;
wherein, the loss function calculates the average difference between all relationship scores output in an epamode and the consistency of corresponding sample pairs;
wherein, the sample is binary to type coincidence degree 0, 1, specifically: when the two sample labels in the sample pair are consistent, the label is 1, and when the two sample labels are inconsistent, the label is 0;
step 5.4, updating each parameter of the AJSARNET by using an optimization method based on the loss function output in the step 5.3;
step 6: testing the trained AJSARNET output in the step 5 based on the support set and the test set output in the step 2 to obtain a network performance test result;
the performance test result specifically comprises the following two contents:
1) and (3) testing loss: calculating the relation score of the input sample pair by the AJSARNET and the average difference of the type consistency of the sample pair by using a loss function;
the sample pairs comprise two AJS signal samples which are respectively from the support set and the test set output in the step 2;
the sample pair type consistency is binary between 0 and 1, specifically: when the two sample labels in the sample pair are consistent, the label is 1, and when the two sample labels are inconsistent, the label is 0;
2) and (3) testing precision: the accuracy of the classification result of the sample in the test set is measured by the average absolute error and the mean square error shown in the formulas (6) and (7) respectively;
wherein MSE, Mean Square Error, is the Mean Square Error; MAE, Mean Absolute Error, is the Mean Absolute Error; m refers to the total number of samples in the test set; n is a radical ofiThe type consistency of the input ith sample pair is obtained;
calculating the relation score of the ith sample pair by using the trained AJSARNET output in the step 5;
step 6, specifically comprising the following substeps:
step 6.1, inputting a sample pair consisting of the support set output in the step 2 and the test concentrated sample into the trained AJSARNET output in the step 5, and outputting a corresponding relation score;
the mode of forming the sample pair specifically includes: each sample in the test set is paired with each sample in the support set;
step 6.2, inputting all the relation scores output in the step 6.1 and the type consistency of the corresponding sample pairs into a calculation formula of a loss function to obtain test loss;
the type consistency of the sample pair is binary 0 and 1, specifically: when the two sample labels in the sample pair are consistent, the label is 1, and when the two sample labels are inconsistent, the label is 0;
step 6.3, based on the relation score of each sample pair output in the step 6.2, calculating MAE and MSE of the test result according to the formulas (6) and (7) to obtain the test precision of the model;
to this end, from step 1 to step 6, an active interference identification method based on an adaptive identification network is completed.
Advantageous effects
The active interference identification method based on the self-adaptive identification network provided by the invention realizes effective self-adaptive active interference identification under the background of low interference-to-noise ratio (JNR), and has the following beneficial effects compared with the prior art:
1. the method establishes a proper AJS simulation data set of a small sample, the data set is divided into four parts, and the proposed model can be easily trained through meta-migration learning; noise is added to the analog data set, so that the robustness or generalization capability of the pre-trained neural network can be improved;
2. the AJSARNET model constructed by the method is light in weight, high-efficiency and low-cost network modules such as a residual error network, an attention mechanism and a meta-transfer learning strategy are integrated, and on the premise of slightly increasing the calculation load, the AJSARNET feature extraction capability on input is improved to a greater extent, so that the AJS recognition accuracy of the model is improved;
3. the method introduces a relation module which dynamically learns the relation score of an input sample pair instead of a predefined fixed measurement, so that the defect of the predefined fixed similarity measurement is overcome, and the AJS recognition accuracy and efficiency of the model are improved;
4. the training process of the method adopts event-based meta-migration learning, allows the model to effectively learn the prior knowledge irrelevant to the task from the training set, and with the help of the prior knowledge and the support set, a trained model can transfer and learn through a small sample, and perfectly allocates a corresponding label for the test sample.
Drawings
FIG. 1 is a flow chart of an active interference identification method based on an adaptive identification network according to the present invention;
fig. 2 is a schematic diagram of an AJSARNet network structure on which the active interference identification method based on the adaptive identification network of the present invention depends.
Detailed Description
The active interference identification method based on the adaptive identification network according to the present invention is described in detail with reference to the accompanying drawings and specific embodiments.
Example 1
The method provided by the invention can be theoretically applied to any small sample learning scene which can be converted into image classification, such as handwriting recognition. In different application scenarios, only the kind of image in the data set needs to be changed.
This embodiment illustrates a specific implementation of the active interference identification method based on the adaptive identification network according to the present invention, and a flow thereof is shown in fig. 1.
The embodiment describes the specific implementation of the active interference identification method based on the adaptive identification network in the 5-way 5-shot scene of identifying six types of AJS signals.
Wherein, six types of AJS signal are specifically:
1) linear Frequency Modulation (LFM), which is a Linear Frequency Modulation signal;
2) Non-Linear Frequency Modulation (NLFM) is a Non-Linear Frequency Modulation (Non-Linear Frequency Modulation);
3) a binary Frequency Shift Keying signal, 2Frequency Shift Keying, abbreviated as 2 FCK;
4) a quaternary Frequency Shift Keying signal, i.e., 4Frequency Shift Keying, abbreviated as 4 FCK;
5) phase encoding, abbreviated PC;
6) a Single Carrier Frequency signal, or Single Carrier Frequency, abbreviated SCF;
firstly, simulating the six AJS types by using six function tools on an MATLAB, and generating 320 different AJS signals by setting different signal parameters such as signal amplitude, carrier frequency, initial phase and the like for each AJS type to generate 1920 AJS signals; then, Gaussian white noise with different amplitudes is added to the AJS signals respectively, so that each AJS type signal is distributed at 16 dry-to-noise ratio levels which take 2dB as a step length and are from-20 dB to 10dB, and each 20 AJS signals are at the same dry-to-noise ratio level;
here, the method adds noise to the analog data set, which can improve the robustness or generalization ability of the pre-trained neural network; the existing AJS identification method often has the problems of poor robustness and low generalization capability.
Secondly, performing short-time Fourier transform on each generated AJS signal to generate a corresponding two-dimensional time-frequency graph, taking the type of the AJS signal as a label of the AJS signal, and generating an AJS simulation data set by using a sample formed by the AJS two-dimensional time-frequency spectrogram and the label of the AJS signal;
here, the method establishes a suitable AJS simulation dataset of small samples, so the proposed model can be easily trained by metastasized learning;
taking all samples with the same AJS type and the same dry-to-noise ratio level in the simulation data set as a group of data, wherein the simulation data set comprises 96 groups of data; as the problem to be solved by the embodiment is the identification of 5-way 5-shot, 5 groups of data are randomly selected from the AJS simulation data set, 5 samples are randomly selected from each group of data to form a support set, and the rest samples form a test set; randomly selecting 5 samples from the remaining 91 groups of data to form a sample set, and forming the remaining samples into a query set;
an AJS adaptive identification network is established, called AJSARNET, and the specific network structure is as follows: cascading the input, embedded module, relationship module, and output as shown in FIG. 2; the embedded module takes four convolutional layers as a main body of the embedded module, and introduces a residual error network and an SE module on the basis;
the AJSARNET model constructed by the method is light in weight, high-efficiency and low-cost network modules such as a residual error network, an attention mechanism and a meta-transfer learning strategy are integrated, and on the premise of slightly increasing the calculation load, the AJSARNET feature extraction capability on input is improved to a greater extent, so that the AJS recognition accuracy of the model is improved; the existing AJS identification method is often difficult to meet the requirements of light weight, high efficiency and high precision at the same time;
the relation module consists of two convolution layers and two full-connection layers, and the last full-connection layer outputs a relation score to determine a final recognition result;
here, the method introduces a relationship module, dynamically learns the relationship score of the input sample pair, and never improves the AJS recognition accuracy and efficiency of the model; the existing AJS recognition method predefines a fixed similarity measurement so that the recognition accuracy and efficiency are to be improved;
then, initializing the network of the constructed AJSARNET (as shown in figure 2), and assigning an initial value to each parameter of the neural network;
next, the number of training rounds is set to 200,000 epicodes, and in each epicode, 5 classes of data, 20 samples per class, will be randomly selected from the AJS simulation dataset.
Wherein, 5 samples in each category are randomly selected as a sample set, and the remaining 15 samples form a query set to form 5 × 5 × 15-375 sample pairs.
Inputting the initialized AJSARNET by each sample pair, outputting a corresponding correlation coefficient, substituting the correlation coefficient output by an epsilon and the type consistency of the sample pairs into a loss function shown in a formula (8), and calculating training loss; then, an Adam optimizer is adopted to update the model parameters based on the training loss, and the training of an epicode is completed(ii) a Wherein the initial learning rate of the Adam optimizer is set to 10-3And the learning rate is reduced by half every 10 ten thousand epidesoses run;
wherein r isi,jRepresenting the relation scores of a sample i in a sample set and a sample j in a query set in the current epamode calculated by AJSARNET; liAnd ljLabels representing sample i and sample j, respectively;
finally, inputting a sample pair formed by a support set and a test set of the AJS simulation data set into the trained AJSARNET, comparing the output relationship fraction with the sample pair type consistency, and respectively calculating the average absolute error (MAE) and the Mean Square Error (MSE) according to the formulas (1) and (2), thereby obtaining the AJS recognition precision of the model and completing the test of the model performance;
the method provides an event-based meta-transfer learning method, which allows a model to effectively learn the prior knowledge which is irrelevant to the task from a training set, and with the help of the prior knowledge and a support set, a trained model can be transferred and learned through a small sample, and a corresponding label is perfectly allocated to a test sample; in the existing AJS recognition method, only a model with low recognition precision can be designed under the condition of insufficient data samples;
therefore, the specific implementation of the active interference identification method based on the adaptive identification network in the 5-way 5-shot scene of identifying the six types of AJS signals is completed.
While the foregoing is directed to the preferred embodiment of the present invention, it is not intended that the invention be limited to the embodiment and the drawings disclosed herein. Equivalents and modifications may be made without departing from the spirit of the disclosure, which is to be considered as within the scope of the invention.
Claims (10)
1. An active interference identification method based on a self-adaptive identification network relies on an AJS self-adaptive identification network model, namely AJSARNET; the AJSARNET is formed by cascading an embedded module and a relation module; the input of the embedded module is the input of AJSARNET, and the embedded module learns the high-dimensional feature expression of the input sample pair and outputs an effective embedded vector; the relation module comprises two convolution layers and two full-connection layers, and the last full-connection layer outputs a relation score to determine a final recognition result; the input of the relation module is the output of the embedded module, namely: a valid embedding vector; the relationship module learns a mapping between valid embedding vectors output by the embedded module and relationship scores of pairs of input samples, characterized in that: the method comprises the following steps:
step 1: constructing an active interference signal simulation data set by using time-frequency transformation;
wherein, active jamming signal is abbreviated as AJS;
the active interference signal simulation data set generated in the step 1 is used for training, testing and evaluating the AJS recognition model and comprises NxSxM samples;
wherein N refers to the type number of the AJS signals in the constructed AJS analog data set; s means that samples contained in each type of AJS signal comprise S different interference-to-noise ratio levels; m means that each class of AJS signals contains M samples at each dry-to-noise ratio level;
step 1, specifically comprising the following substeps:
step 1.1, constructing an AJS simulation data set, which specifically comprises the following steps: according to the requirement of the AJS recognition task, simulating N multiplied by S multiplied by M AJS signals to generate an AJS signal set;
step 1.2, performing time-frequency transformation on each AJS signal in the AJS signal set output in the step 1.1 to generate a corresponding two-dimensional time-frequency spectrogram, taking the type of the AJS signal as a label of the AJS signal, and forming an active interference signal simulation data set by using the AJS two-dimensional time-frequency spectrogram and the label of the AJS signal;
step 2: dividing the active interference signal simulation data set output in the step 1 according to the set identification problem of the C-way K-shot to obtain a sample set, a query set, a support set and a test set of the AJS identification model;
the sample set and the query set jointly form a training set for training the AJS recognition model; the supporting set and the testing set are used for testing the AJS recognition model, the supporting set contains C-class data, and each class of data consists of K samples;
the method specifically comprises the following substeps of dividing the active interference signal analog data set output in the step 1:
step 2.1, randomly selecting a C-type AJS signal in the active interference signal simulation data set output in the step 1, and dividing the C-type AJS signal to form a support set and a test set;
step 2.2, removing samples belonging to the support set or the test set generated in the step 2.1 from the active interference signal simulation data set output in the step 1, and forming a training set by the remaining samples;
step 2.3, dividing the training set output in the step 2.2 to form a sample set and a query set;
and step 3: according to the requirements of the AJS recognition task, building AJSARNET;
the input of the AJSARNET comprises a sample pair formed by two AJS two-dimensional time-frequency spectrograms; the output of the AJSARNET is the relation score of the input sample pair;
and 4, step 4: initializing the AJSARNET established in the step 3;
initializing AJSARNET, specifically assigning an initial value to each parameter of a neural network;
and 5: training the initialized AJSARNET output in the step 4 based on the sample set and the query set output in the step 2 to obtain the trained AJSARNET;
specifically, setting L element training iteration units epicode for the initialized AJSARNET training output in the step 4;
for the set C-way K-shot recognition task, setting an epsilon to send all samples of the selected meta-training set into a network to be trained, and completing a forward calculation and back propagation process for each sample, specifically: selecting C classes in the training set formed by the sample set and the query set output in step 2, training a primary model by using all samples of the selected C class data, and updating the primary modelA parameter; selecting other C class training models for the next epicode; for the sample set containing NxS-C class data and the query set output in step 2, each epoch contains
An epicode;
when the NxS-C is not evenly divisible by C, the last epadiode of the current epoch represents that the type of the untrained data is less than the type C, and at the moment, the C type needs to be randomly selected and supplemented from the trained data to train the last epadiode;
specifically, the training process of each epicode includes the following sub-steps:
step 5.1, respectively taking out C-type AJS signals with the same type from the sample set and the query set output in the step 2 to generate a sample pair;
the sample set output in the step 2 and the query set respectively contain K samples and M-K samples for each type of AJS signals;
step 5.2, sequentially inputting the sample pairs output in the step 5.1 into AJSARNET to be trained according to the types of the sample sets of the sample pairs for calculation to obtain relationship scores;
wherein, AJSARNET to be trained specifically means: if the epsilon is the first epsilon of the model training stage, the initialized AJSARNET output in the step 4 is obtained, otherwise, the initialized AJSARNET output by the last epsilon is obtained;
step 5.3, substituting the relation score output in the step 5.2 into a loss function formula to calculate the training loss;
wherein, the loss function calculates the average difference between all relationship scores output in an epamode and the consistency of corresponding sample pairs;
wherein, the sample is binary to type coincidence degree 0, 1, specifically: when the two sample labels in the sample pair are consistent, the label is 1, and when the two sample labels are inconsistent, the label is 0;
step 5.4, updating each parameter of the AJSARNET by using an optimization method based on the loss function output in the step 5.3;
step 6: testing the trained AJSARNET output in the step 5 based on the support set and the test set output in the step 2 to obtain a network performance test result;
step 6, specifically comprising the following substeps:
step 6.1, inputting a sample pair consisting of the support set output in the step 2 and the test concentrated sample into the trained AJSARNET output in the step 5, and outputting a corresponding relation score;
step 6.2, inputting all the relation scores output in the step 6.1 and the type consistency of the corresponding sample pairs into a calculation formula of a loss function to obtain test loss;
and 6.3, calculating the MAE and the MSE of the test result based on the relation score of each sample pair output in the step 6.2 to obtain the test precision of the model.
2. The active interference identification method of claim 1, characterized in that: each sample in step 1 consists of two parts: 1) a two-dimensional time-frequency spectrogram is generated after the AJS signal is subjected to time-frequency transformation; 2) the tag of the AJS signal is the true value of the AJS recognition result of the signal.
3. The active interference identification method of claim 1, characterized in that: the dry-to-noise ratio refers to the energy ratio of signal to noise, i.e. Jamming-to-noise, abbreviated as JNR, and is defined by formula (1):
wherein, PjAnd PnInterference power and noise power, respectively.
4. The active interference identification method of claim 1, characterized in that: step 1.1, comprising the following substeps:
step 1.1A, selecting N different modulation modes, and determining N types of the AJS signals;
step 1.1B, according to the N types of the AJS signals determined in the step 1.1A, setting different signal parameters for each type of the AJS signals by using N different modulations, respectively generating S multiplied by M AJS analog signals, and generating an original AJS signal set;
step 1.1C, S dry-to-noise ratio levels are set, and white noise with S amplitudes is added to the AJS signals in the original AJS signal set generated in step 1.1B according to formula (1) to generate an AJS signal set; step 1.1C, specifically: for every M signal samples of each type of AJS signal, white noise of the same amplitude is added, each type of AJS signal containing S uniform different levels of interference to noise ratio.
5. The active interference identification method of claim 1, characterized in that: in the step 2, the problem of C-way K-shot recognition specifically refers to a classification task constructed in meta-learning; the class C data specifically refers to a set formed by samples of which the types and the dry-to-noise ratio levels of the AJS signals are the same in the AJS analog data set output in the step 1;
in step 2.1, the specific division method is as follows: for each type of the C type AJS signals, randomly selecting K samples to generate a support set, and generating a test set by the rest samples; the specific dividing method in the step 2.3 comprises the following steps: and (3) randomly selecting K samples for each class in the training set output in the step 2.2 to generate a sample set, and generating a query set by using the rest samples in the training set.
6. The active interference identification method of claim 1, characterized in that: in step 3, the relationship score is a number from 0 to 1, and is used for measuring the similarity of two samples in the sample pair: the larger the relation score is, the more similar the two samples are, and when the relation score is larger than a certain threshold value, the two samples can be judged to be the AJS signals of the same type.
7. The active interference identification method of claim 1, characterized in that: step 5.1, specifically: respectively with SijAnd QijDenotes the j-th sample, S, of the i-th signal of the class C AJS signal taken from the sample set and the query setijAnd QijThe expression of (c) is specifically shown in formula (2) and formula (3);
Sij＝(xij，li)，i＝1，2，…，C，j＝1，2，…，K (2)
Qij＝(yij，li)，i＝1，2，…，C，j＝1，2，…，M-K (3)
wherein x and y respectively represent a two-dimensional time-frequency spectrogram of a sample set and a sample of a query set, and l represents a label of the sample;
the generated sample pair is specifically shown as formula (4);
Pijmn＝(Sij，Qmn) (4)
wherein, i is 1, 2, …, C, j is 1, 2, …, K; m-1, 2, …, C, n-1, 2, …, M-K; i.e. each sample in the sample set taken in step 6.1, forms a sample pair with each sample in the query set taken in step 6.1, and each epamode generates C × K × C × (M-K) sample pairs in common.
8. The active interference identification method of claim 1, characterized in that: step 5.2, specifically comprising the following substeps:
step 5.2.1 groups the sample pairs output in step 5.2 to obtain a sample pair group Timn；
Wherein, i is 1, 2, …, C, M is 1, 2, …, C, n is 1, 2, …, M-K, i.e. the generated sample pair group has C × (M-K) × C;
grouping the sample pairs specifically includes: sample a query set by QmnAnd samples S of a certain type of sample setijK sample pairs of 1, 2, …, K are used as a sample pair group TimnAs shown in formula (5);
Timn＝{Pijmn|j＝1，2，…，K}＝{(Sij，Qmn)|j＝1，2，…，K} (5)
step 5.2.2 pairs the samples output in step 5.2.1 against group TimnSequentially inputting the embedded modules of AJSARNET to be trained for calculation to obtain the total embedded vector of each sample pair group;
wherein, i is 1, 2, …, C, M is 1, 2, …, C, n is 1, 2, …, M-K;
specifically, the embedded module of the AJSARNET to be trained can pair the input samples into a group TimnOutputs a valid embedded vector, i.e. a sample pair group TimnGenerating K effective embedding vectors, wherein the sum of the K effective embedding vectors is TimnThe total embedded vector of (1);
step 5.2.3, sequentially inputting the total embedded vector output in the step 5.2.2 into a relation module of AJSARNET to be trained to obtain a relation score of each sample pair group;
specifically, the relationship score of each sample to the group describes the similarity between a certain query set sample and a certain type of sample set sample.
9. The active interference identification method of claim 1, characterized in that: in step 6, the performance test result specifically includes the following two parts:
1) and (3) testing loss: calculating the relation score of the input sample pair by the AJSARNET and the average difference of the type consistency of the sample pair by using a loss function;
the sample pairs comprise two AJS signal samples which are respectively from the support set and the test set output in the step 2;
the sample pair type consistency is binary between 0 and 1, specifically: when the two sample labels in the sample pair are consistent, the label is 1, and when the two sample labels are inconsistent, the label is 0;
2) and (3) testing precision: the accuracy of the classification result of the sample in the test set is measured by the average absolute error and the mean square error shown in the formulas (6) and (7) respectively;
wherein MSE, Mean Square Error, is the Mean Square Error; MAE, Mean Absolute Error, is the Mean Absolute Error; m refers to the total number of samples in the test set; n is a radical ofiThe type consistency of the input ith sample pair is obtained;
and (5) calculating a relation score of the ith sample pair by using the trained AJSARNET output in the step 5.
10. The active interference identification method of claim 1, characterized in that: in step 6.1, the mode of forming a sample pair specifically includes: each sample in the test set is paired with each sample in the support set;
in step 6.2, the type consistency of the sample pair is binary between 0 and 1, specifically: the two sample labels in a sample pair are 1 if they are identical, and 0 if they are not identical.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111318774.5A CN114201987B (en) | 2021-11-09 | 2021-11-09 | Active interference identification method based on self-adaptive identification network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111318774.5A CN114201987B (en) | 2021-11-09 | 2021-11-09 | Active interference identification method based on self-adaptive identification network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114201987A true CN114201987A (en) | 2022-03-18 |
| CN114201987B CN114201987B (en) | 2024-07-09 |
Family
ID=80647104
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111318774.5A Active CN114201987B (en) | 2021-11-09 | 2021-11-09 | Active interference identification method based on self-adaptive identification network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114201987B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114818777A (en) * | 2022-03-18 | 2022-07-29 | 北京遥感设备研究所 | Training method and device for active angle deception jamming recognition model |
| CN115001609A (en) * | 2022-04-29 | 2022-09-02 | 国网上海市电力公司 | Electric power wireless private network interference type identification method considering interference power measured value |
| CN116482618A (en) * | 2023-06-21 | 2023-07-25 | 西安电子科技大学 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110764064A (en) * | 2019-11-08 | 2020-02-07 | 哈尔滨工业大学 | Radar interference signal identification method based on deep convolutional neural network integration |
| US20200285896A1 (en) * | 2019-03-09 | 2020-09-10 | Tongji University | Method for person re-identification based on deep model with multi-loss fusion training strategy |
| CN112731309A (en) * | 2021-01-06 | 2021-04-30 | 哈尔滨工程大学 | Active interference identification method based on bilinear efficient neural network |
| CN113420593A (en) * | 2021-05-19 | 2021-09-21 | 西安电子科技大学 | Small sample SAR automatic target recognition method based on hybrid inference network |
-
2021
- 2021-11-09 CN CN202111318774.5A patent/CN114201987B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200285896A1 (en) * | 2019-03-09 | 2020-09-10 | Tongji University | Method for person re-identification based on deep model with multi-loss fusion training strategy |
| CN110764064A (en) * | 2019-11-08 | 2020-02-07 | 哈尔滨工业大学 | Radar interference signal identification method based on deep convolutional neural network integration |
| CN112731309A (en) * | 2021-01-06 | 2021-04-30 | 哈尔滨工程大学 | Active interference identification method based on bilinear efficient neural network |
| CN113420593A (en) * | 2021-05-19 | 2021-09-21 | 西安电子科技大学 | Small sample SAR automatic target recognition method based on hybrid inference network |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114818777A (en) * | 2022-03-18 | 2022-07-29 | 北京遥感设备研究所 | Training method and device for active angle deception jamming recognition model |
| CN114818777B (en) * | 2022-03-18 | 2023-07-21 | 北京遥感设备研究所 | Training method and device for active angle deception jamming recognition model |
| CN115001609A (en) * | 2022-04-29 | 2022-09-02 | 国网上海市电力公司 | Electric power wireless private network interference type identification method considering interference power measured value |
| CN116482618A (en) * | 2023-06-21 | 2023-07-25 | 西安电子科技大学 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
| CN116482618B (en) * | 2023-06-21 | 2023-09-19 | 西安电子科技大学 | Radar active interference identification method based on multi-loss characteristic self-calibration network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114201987B (en) | 2024-07-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | A fault diagnosis method for wind turbines gearbox based on adaptive loss weighted meta-ResNet under noisy labels | |
| CN114201987B (en) | Active interference identification method based on self-adaptive identification network | |
| CN109993280B (en) | Underwater sound source positioning method based on deep learning | |
| CN109597043B (en) | Radar signal identification method based on quantum particle swarm convolutional neural network | |
| CN110222748B (en) | OFDM radar signal identification method based on 1D-CNN multi-domain feature fusion | |
| CN109635245A (en) | A kind of robust width learning system | |
| Liang et al. | A transfer learning approach for compressed sensing in 6G-IoT | |
| CN112036239A (en) | Radar signal working mode identification method and system based on deep learning network | |
| CN112733447A (en) | Underwater sound source positioning method and system based on domain adaptive network | |
| CN113780242A (en) | Cross-scene underwater sound target classification method based on model transfer learning | |
| CN113779873A (en) | Construction method and application of TDECQ measurement model | |
| CN115061103A (en) | Multifunctional radar function identification method based on encoder-decoder | |
| CN115659254A (en) | Power quality disturbance analysis method for power distribution network with bimodal feature fusion | |
| CN113111786A (en) | Underwater target identification method based on small sample training image convolutional network | |
| CN114879185A (en) | Radar target intelligent identification method based on task experience migration | |
| CN117216517A (en) | Radiation source individual identification method based on contrast generation sample enhancement | |
| CN115238748B (en) | Modulation identification method based on Transformer and decision fusion | |
| CN115422977A (en) | Radar radiation source signal identification method based on CNN-BLS network | |
| CN113343924B (en) | Modulation signal identification method based on cyclic spectrum characteristics and generation countermeasure network | |
| Pan et al. | Specific radar emitter identification using 1D-CBAM-ResNet | |
| CN115598714A (en) | Time-space coupling neural network-based ground penetrating radar electromagnetic wave impedance inversion method | |
| CN113343796B (en) | Knowledge distillation-based radar signal modulation mode identification method | |
| CN115329821A (en) | Ship noise identification method based on pairing coding network and comparison learning | |
| CN113189571B (en) | Sound source passive ranging method based on tone feature extraction and deep learning | |
| CN111931412A (en) | Underwater target noise LOFAR spectrogram simulation method based on generative countermeasure 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 |