CN114970638A - Radar radiation source individual open set identification method and system - Google Patents

Radar radiation source individual open set identification method and system Download PDF

Info

Publication number
CN114970638A
CN114970638A CN202210652213.7A CN202210652213A CN114970638A CN 114970638 A CN114970638 A CN 114970638A CN 202210652213 A CN202210652213 A CN 202210652213A CN 114970638 A CN114970638 A CN 114970638A
Authority
CN
China
Prior art keywords
signal
radiation source
radar
identification
identified
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210652213.7A
Other languages
Chinese (zh)
Inventor
韩啸
陈世文
陈蒙
杨锦程
东锦鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Information Engineering University of PLA Strategic Support Force
Original Assignee
Information Engineering University of PLA Strategic Support Force
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Information Engineering University of PLA Strategic Support Force filed Critical Information Engineering University of PLA Strategic Support Force
Priority to CN202210652213.7A priority Critical patent/CN114970638A/en
Publication of CN114970638A publication Critical patent/CN114970638A/en
Pending legal-status Critical Current

Links

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
    • G06F2218/12—Classification; Matching
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • 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
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • G06N3/08—Learning methods
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

The invention belongs to the technical field of radar radiation source identification, and particularly relates to a radar radiation source individual open set identification method and a system, wherein a signal identification model is constructed, and a marked signal sample in a radar signal database is utilized to train the signal identification model, wherein a transient sequence of a radar signal is adopted as the network input of the signal identification model, the input signal characteristics are extracted through a signal identification model network, and the extracted signal characteristics are classified and identified; the method comprises the steps of inputting radar radiation source individual signals to be identified, intercepted by a receiver, as signal identification models, identifying the radar radiation source individual signals to be identified by utilizing the trained signal identification models, if the radar radiation source individual signals are identified as known classes, outputting radiation source names of the known classes, and if the radar radiation source individual signals are identified as unknown classes, marking the signals as unknown classes, adding the signals into a radar signal database, and identifying and marking the signals manually. The method has the advantages of good generalization, high identification speed, capability of classifying known radiation sources and capability of detecting new unknown radiation sources, and convenience for actual scene application.

Description

Radar radiation source individual open set identification method and system
Technical Field
The invention belongs to the technical field of radar radiation source identification, and particularly relates to a radar radiation source individual open set identification method and system.
Background
The current electromagnetic environment presents a trend of increasing complexity, the signal density is increasing, and aliasing in time domain and frequency domain is serious. Radars play more and more important roles in life, and particularly with the accelerated popularization and application of technologies such as automatic driving, radar imaging, radar detection and the like, radar signals become important components in electromagnetic environments. In the face of a complex electromagnetic environment, the traditional radiation source identification method based on the pulse description words has performance bottlenecks, and different individual radars with the same model cannot be distinguished. Meanwhile, due to the development of software radio technology, radar parameters are variable, and the radiation source is more difficult to identify through traditional parameters.
Under the background, many existing methods extract features capable of reflecting hardware differences of different transmitters by analyzing radar signals, so that individual identification of radar radiation sources is realized. These features, also referred to as fingerprint features, are due to process gaps in the components of the radiation source hardware. Similar to human fingerprints, radiation source fingerprint characteristics are also not easily modified unless hardware changes occur. Therefore, the individual identification of the radiation source through fingerprint characteristics is a good idea. Some studies extract features that distinguish particular radiation sources based on expert knowledge, such as intra-pulse unintentional phase modulation features, unintentional amplitude modulation features, etc., and then classify the features in conjunction with a classifier. The method has the advantages that the calculated amount of the feature calculation is small, and the speed is high; the disadvantages are that a strong expert knowledge background is needed, the feature design process is complicated, and the generalization of the features is not strong. With the development of artificial intelligence technology, more and more researchers use deep learning instead of manual extraction of fingerprint features. Some methods need to perform complex preprocessing on signals and then input the signals into a neural network, and the methods have long signal preprocessing time and low recognition efficiency during recognition. In other methods, original signal data are directly used as network input, but the characteristics of radar signal fingerprint characteristics are not combined, so that a certain degree of input redundancy exists, and greater computing pressure is brought to a subsequent characteristic extraction network. In addition, the existing method has the characteristic of openness in consideration of the real electromagnetic environment. Especially based on discriminant functions, machine learning and deep learning, if the identified signal comes from a new radiation source completely without prior information, it will be mistaken for a known class. This is a very large hole for spectrum monitoring systems that need to detect illegal users.
Disclosure of Invention
Therefore, the radar radiation source individual open set identification method and system provided by the invention have the advantages of good generalization, high identification speed, capability of classifying known radiation sources and new unknown radiation source detection capability, and convenience for actual scene application.
According to the design scheme provided by the invention, a radar radiation source individual open set identification method is provided, which comprises the following contents:
constructing a signal recognition model, and training the signal recognition model by using marked signal samples in a radar signal database, wherein a transient sequence of a radar signal is used as a signal recognition model network input, and the input signal characteristics are extracted through the signal recognition model network and are classified and recognized;
the method comprises the steps of inputting radar radiation source individual signals to be identified, intercepted by a receiver, as signal identification models, identifying the radar radiation source individual signals to be identified by utilizing the trained signal identification models, if the radar radiation source individual signals are identified as known classes, outputting radiation source names of the known classes, and if the radar radiation source individual signals are identified as unknown classes, marking the signals as unknown classes, adding the unknown classes into a radar signal database, and identifying and marking manually.
As the radar radiation source individual open set identification method, further, a multi-scale ResNet network is adopted in the signal identification model, and the multi-scale ResNet network comprises the following components: the system comprises an SK module for extracting input multi-channel signal characteristics, a characteristic fusion module for performing fusion processing on the multi-channel signal characteristics, an average pooling layer for performing average pooling operation on the fusion characteristics to obtain characteristic vectors, a full connection layer for performing dimension processing on the characteristic vectors, a softmax layer for performing classification operation on the characteristic vectors after the dimension processing to obtain channel coefficients, and an output layer for fusing the channel coefficients and the corresponding channel signal characteristics extracted by the SK module and determining a final classification recognition result.
As the radar radiation source individual open set identification method, further, the SK module adopts a one-dimensional residual convolution structure, and the one-dimensional residual convolution structure includes a plurality of different convolution kernel paths for extracting signal characteristics through convolution operation.
As the radar radiation source individual open set identification method, further, in the training of the signal identification model, firstly, the marked signal sample is used for training the signal identification model to obtain a closed set signal identification network, and an output characteristic activation vector is obtained through a full connection layer; then, retaining the correctly classified activation vectors, and forming an activation vector set according to different classes; then, calculating a distance set from the activation vector to the activation vector mean under each category element in the activation vector set; and finally, sequencing the distance values in the distance set, selecting the largest epsilon distance values of each type after sequencing, and obtaining an extreme value theoretical model corresponding to the type of the marked signal sample by utilizing an extreme value theory and fitting a probability density function of the mean value of the activated vectors of each type of the marked signal sample in the feature space to the extreme value.
As the radar radiation source individual open set identification method, further, the target loss function of the signal identification model training is expressed as:
Figure BDA0003688143380000021
where N is the number of marked signal samples, F (x) i ) A feature vector representing a fully connected layer output;
Figure BDA0003688143380000022
and b i Representing full connection layer weights and offsets;
Figure BDA0003688143380000023
and the signal sample of the ith class is represented in the feature space center of the class to which the signal sample belongs, and lambda is a hyperparameter for adjusting the loss weight.
As the radar radiation source individual open set identification method, further, the Euclidean distance is adopted to measure the distance from the activation vector to the mean value of the activation vector.
As the radar radiation source individual open set identification method, further, in the identification of the radar radiation source individual signal to be identified by using a trained signal identification model, firstly, the radar radiation source individual signal to be identified is used as the model input to obtain the activation vector of the input signal, and the distance between the activation vector and the mean value of the activation vector of each marked signal sample category is calculated; then, the mean value is brought into a corresponding extreme value theoretical model, a reliability score is obtained through the extreme value theoretical model, and the reliability score is used for adjusting the activation vector value; and finally, identifying the radiation source category by normalizing the probabilities of the various types of input signals.
As the radar radiation source individual open set identification method, further, the reliability score calculation formula is expressed as follows: .
Figure BDA0003688143380000031
Wherein K, s (j), v i Respectively representing the number of the known radiation source types, the serial number of the classification index of the activation vector with the size of j, the activation vector of the ith identification sample, and theta s(j) 、λ s(j) 、κ s(j) Three parameters of the s (j) th extremum model are respectively represented.
As the radar radiation source individual open set identification method, further, various signal probability formulas are expressed as follows:
Figure BDA0003688143380000032
wherein x is i Indicating to be identifiedThe signal(s) of the signal(s),
Figure BDA0003688143380000033
which is indicative of the class of the signal to be identified,
Figure BDA0003688143380000034
indicating the adjusted activation vector value.
Further, the present invention also provides a radar radiation source individual open set identification system, comprising: a model training module and a signal recognition module, wherein,
the model training module is used for training the signal recognition model by constructing the signal recognition model and utilizing the marked signal samples in the radar signal database, wherein the transient sequence of the radar signal is used as the network input of the signal recognition model, the input signal characteristics are extracted through the signal recognition model network, and the extracted signal characteristics are classified and recognized;
and the signal identification module is used for inputting the radar radiation source individual signals to be identified intercepted by the receiver as a signal identification model, identifying the radar radiation source individual signals to be identified by using the trained signal identification model, if the radar radiation source individual signals are identified as a known class, outputting the name of the radiation source belonging to the known class, and if the radar radiation source individual signals are identified as an unknown class, marking the signals as the unknown class and adding the unknown class into a radar signal database so as to identify and mark the signals manually.
The invention has the beneficial effects that:
the invention can extract the fingerprint characteristics of the radiation source, classify the known radiation source equipment according to the fingerprint characteristics and identify the unknown radiation source; the radiation source fingerprint characteristics are extracted by adopting a one-dimensional convolution neural network, the expert knowledge is not depended on, and the generalization performance is better; by utilizing the SKResNet with a residual network structure, the training degradation and different receptive fields of a multi-scale convolution kernel are prevented, the receptive fields can be dynamically adjusted according to the input and the loss, and the characteristics with better representation can be extracted; in the preprocessing process, the transient sequence of the radar signal is extracted, the calculation process is efficient, and compared with the method that the complete pulse signal is directly input into a neural network, the calculation and storage resources are saved; and considering that a real signal environment is open, the SKResNet network is improved based on the OpenMax method, so that the SKResNet network has the open set identification capability, and the unknown radiation sources can be identified while the known radiation sources are correctly classified. And the method is provided with the CenterLoss for training according to the OpenMax principle, so that the performance of the method is further improved.
Description of the drawings:
FIG. 1 is a schematic flow chart of an individual open set identification of a radar radiation source in the embodiment;
FIG. 2 is a schematic diagram of signal preprocessing flow in the embodiment;
FIG. 3 is a structural schematic of an SK module in the embodiment;
FIG. 4 is a schematic diagram of an SKResNet structure in the embodiment;
FIG. 5 is a schematic diagram showing the comparison of the sample distribution in the feature space before and after the CenterLoss training is added in the example;
FIG. 6 is a schematic diagram showing the test results of the present embodiment.
The specific implementation mode is as follows:
in order to make the objects, technical solutions and advantages of the present invention clearer and more obvious, the present invention is further described in detail below with reference to the accompanying drawings and technical solutions.
The embodiment of the invention provides a radar radiation source individual open set identification method, which comprises the following contents:
constructing a signal recognition model, and training the signal recognition model by using marked signal samples in a radar signal database, wherein a transient sequence of a radar signal is used as a signal recognition model network input, and the input signal characteristics are extracted through the signal recognition model network and are classified and recognized;
the method comprises the steps of inputting radar radiation source individual signals to be identified, intercepted by a receiver, as signal identification models, identifying the radar radiation source individual signals to be identified by utilizing the trained signal identification models, if the radar radiation source individual signals are identified as known classes, outputting radiation source names of the known classes, and if the radar radiation source individual signals are identified as unknown classes, marking the signals as unknown classes, adding the unknown classes into a radar signal database, and identifying and marking manually.
Referring to fig. 1, it is assumed that there are K known similar radar radiation sources individually, where similar refers to a radiation source that cannot be identified by conventional PDW and intra-pulse modulation parameters, etc., but can only be distinguished by individual identification. According to the scheme, only the marked radar radiation source individual signal D in the database is used in the model training stage L ＝{(x i ,y i ),i∈[1,N L ],y i ∈[1,K]Training, y denotes the radiation source number, and i is the signal sample number. Training to obtain a recognition model, and intercepting individual signals D of radar radiation sources by a receiver in a recognition stage R ＝{(x i ,y i ),i∈[1,N R ],y i ∈[1,M]}, total M classes, where M is K Class of known radiation sources, M U A class that belongs to unknown radiation sources. Identifying the intercepted signal through a trained model and an open set identification algorithm, and further giving the name of a known radiation source to which the identification belongs if the intercepted signal is identified as a known type; if the new radiation source is identified as an unknown new radiation source, the new radiation source is marked as an unknown class, and the new radiation source is added into a radar database to wait for subsequent manual marking through other methods, so that the generalization is good, and meanwhile, the identification and detection capabilities of the known radiation source and the unknown radiation source are realized, and the application in an actual scene is facilitated.
The purpose of signal pre-processing is to normalize the signal to fit the input requirements of a one-dimensional network. Common radar signal types include pulse signals and continuous wave signals, and in the scheme, the radar signals which are mainly concerned are the pulse signals. The radiation source emits a radar signal and the transient process of the pulse generation is considered as a transient state of the radar signal. The transient characteristics are influenced by the structure of an emitting system and hardware, and the transient characteristics of different radiation source individuals are different, so that the transient characteristics can be used as fingerprint characteristics for distinguishing the individuals.
Since the input of the one-dimensional neural network is generally a fixed length, the one-dimensional neural network is used for processing signals, and the input signals need to be preprocessed to adapt to the input requirements. The common radiation source identification method based on deep learning directly inputs original signal data into a network, but fingerprint characteristics of a time domain radar signal are mainly concentrated in an unstable state when a rising edge changes, namely a transient part of the rising edge. If a complete pulse is used as an input, relatively large input redundancy is generated, the calculation and storage burden is increased, and the model inference time is increased. Therefore, in the scheme, only the transient part of the radar signal is extracted as the network input, and the preprocessing process can be designed as follows:
step 1, orthogonally sampling an intercepted radar signal to obtain a discrete signal; or subjecting the sampled radar signal to Hilbert transform:
z(n)＝a(n)+jb(n),1≤n≤N (1)
step 2 calculate z (n) fourth order moment:
Figure BDA0003688143380000051
step 3, calculating the energy track of the signal:
Figure BDA0003688143380000052
step 4, maximum and minimum normalization of energy tracks:
Figure BDA0003688143380000053
step 5, intercepting q (N) by using a threshold value to obtain N t Index pairs of start and stop sequences of the transient signals, as shown in fig. 2:
Figure BDA0003688143380000054
assuming that the input length of the network is fixed to L, the instantaneous sequence is obtained from each pair of sequence indices:
Figure BDA0003688143380000055
Figure BDA0003688143380000061
the expression is rounded down, the expression above means that the original data sequence z (n) is divided into a transient sequence with length L, and each transient sequence is centered on the index coordinate of the transient signal and extends forward and backward by the same length (or differs by 1 point).
Step 6 pairs of s i And normalizing the data of the real part and the imaginary part for model training and identification.
Further, in the embodiment of the present disclosure, the signal identification model adopts a multi-scale ResNet network, and the multi-scale ResNet network includes: the system comprises an SK module for extracting input multi-channel signal characteristics, a characteristic fusion module for performing fusion processing on the multi-channel signal characteristics, an average pooling layer for performing average pooling operation on the fusion characteristics to obtain characteristic vectors, a full connection layer for performing dimension processing on the characteristic vectors, a softmax layer for performing classification operation on the characteristic vectors after the dimension processing to obtain channel coefficients, and an output layer for fusing the channel coefficients and the corresponding channel signal characteristics extracted by the SK module and determining a final classification recognition result. Further, the SK module employs a one-dimensional residual convolution structure, and a plurality of different convolution kernel paths for extracting signal features through convolution operations are included in the one-dimensional residual convolution structure.
The generation process of the radar signal is complex and has non-stationary characteristics. Classical radar structures include oscillators, amplifiers, filters, transistors, etc. Hardware differences between different devices result in unintentional modulation, i.e., fingerprint characteristics. These features are very subtle and, therefore, a strong expression performance is required for the feature extraction network. Inspired by the working mechanism of visual cortex, in a feature extraction network which is expected to be designed by a plurality of methods, neurons can have different perception ranges so as to improve the expression capability of the network. The one-dimensional neural network performs one-dimensional convolution operation on the signals, the one-dimensional convolution operation is similar to filtering, one-dimensional convolution kernels with different sizes correspond to the sizes of different filter coefficients, the signal analysis capability of differentiated resolution is achieved, and the method is equivalent to learning signal characteristics from different scale views. The residual error network structure can effectively avoid the problem of gradient explosion or disappearance of the neural network along with the deepening of the depth. When the identification task is complex and the complexity of the model needs to be increased, the mode of residual error connection has great advantage. The Multi-Scale ResNet (MSResNet) can be applied to vibration signal fault detection, and the influence of features extracted by convolution kernels with different scales on a final identification result is different, specifically, the feature of which resolution is set with higher weight and should be automatically adjusted according to the characteristics of an input signal. But the MSResNet directly splices signal characteristics of different scales, and the range of the receptive field is fixed. Therefore, inspired by selectable core size networks (sknets), a multiscale ResNet Network (SKResNet) with a variable receptive field can be designed.
In the embodiment of the scheme, the SKResNet network model is mainly composed of an SK module, and the key of the SK module is that the network dynamically adjusts the weights of different convolution kernel channel characteristics. As shown in fig. 3, the SK module may be designed to have three paths, and the input features of the SK module sequentially pass through one-dimensional residual convolution modules (ResBlock) with sizes of 1 × 3, 1 × 5, and 1 × 7, so as to obtain features with different receptive fields. And carrying out zero padding operation with proper size during convolution operation so that the length of the feature vector is not changed by the convolution operation. ResBlock employs a generic residual network architecture, the output of which is denoted F i out ：
F i out ＝F in +f(F in ,W i ),i＝3,5,7 (7)
F in Indicates the input of ResBlock, W i The parameter value representing the ith ResBlock.
Features from different branches are then fused to obtain fused features.
Figure BDA0003688143380000071
For the fusion feature F fused Carrying out average pooling operation to obtain characteristic quantity V crossing different channels 0 . Using a fully connected network, pairAnd after the dimension of the F is reduced to C', the dimension is increased to C to obtain three feature vectors with the same size, so that the feature selection channel has better selection performance. For the three eigenvectors, a softmax operation is performed in the channel dimension, resulting in coefficient values for the three eigenpaths at each channel.
Figure BDA0003688143380000072
The coefficient value and the original characteristic F i out Multiplying and then adding and fusing to obtain the selected characteristic Y out ：
Figure BDA0003688143380000073
And in the training process, the full-connection layer on the selected path adjusts the weights of different branches according to the model loss. If no feature selection path exists, the branches with different convolution kernel sizes are endowed with the same weight. The model consists of mainly 3 skblocks and two Fully Connected Layers (FCs), as shown in fig. 4. The output of the first FC is the characteristic, and the full connection layer is denoted as F P . In the embodiment of the present invention, F can be replaced by P The output dimension is set to 2 to facilitate visualization of the distribution of features. The output dimension of the second FC is equal to the number K of known radiation sources.
Further, in the embodiment of the present disclosure, in the training of the signal recognition model, firstly, the signal recognition model is trained by using the marked signal sample to obtain a signal recognition network of a closed set, and an output characteristic activation vector is obtained through a full connection layer; then, retaining the correctly classified activation vectors, and forming an activation vector set according to different classes; then, calculating a distance set from the activation vector to the activation vector mean under each category element in the activation vector set; and finally, arranging the distance values in the distance set in sequence, selecting the largest epsilon distance values of each type after sequencing, and obtaining an extreme value theoretical model corresponding to the type of the marked signal sample by utilizing an extreme value theory and fitting a probability density function of the size of the mean value distance extreme value of the activated vector of each type of the marked signal sample in the feature space. Further, the Euclidean distance is used for measuring the distance from the activation vector to the mean value of the activation vector.
Extreme Value Theory (EVT) is often used to describe an arbitrary set of independent and identically distributed random variables x that describe the probability of an Extreme event occurring 1 ,x 2 ,...,x N All satisfy Fisher-Tippett's theorem:
let m equal max { x 1 ,x 2 ,...,x N If there is a set of real pairs (a) n ,b n ) Satisfy any of a n Is greater than 0 and
Figure BDA0003688143380000074
and satisfies:
Figure BDA0003688143380000081
when F (x) is a non-degenerate distribution function, F (x) must belong to the Gumbel family, the Frechet family or the inverse Weibull family of distributions. That is, the maximum value of a group of independent random variables with the same distribution is converged to one of the three distributions according to probability after proper normalization. The three distribution families are actually three cases of generalized extreme value distribution xi > 0, xi ═ 0 and xi < 0, and the cumulative distribution function is as follows:
Figure BDA0003688143380000082
when xi < 0 in the above formula, the distribution form is inverse Weibull distribution form, and the cumulative probability distribution function is equivalently written as:
Figure BDA0003688143380000083
in the above formula, σ, ζ and θ are greater than 0.
The traditional closed set identification network normalizes the identification scores through a softmax layer to obtain the identification probability of each type. However, in the process of identifying the open set scene, due to the influence of the added unknown class, the probability sum of the original class to be classified is no longer 1. OpenMax improves the output of the softmax layer of the traditional closed set identification network, so that the output dimension is increased by one dimension, and the probability that the identification class belongs to unknown class is given. It uses extremum theory to model the distance between the feature vectors (also called activation vectors) of a classification network. During identification, according to the fitted EVT distribution in the identification process, the reliability of the original identification score is reevaluated and corrected, and finally open set identification is carried out.
In the embodiment of the scheme, SKResNet is improved based on OpenMax to form Open-SKResNet, so that the Open-SKResNet has the Open set identification capability. The training algorithm can be designed as follows:
Figure BDA0003688143380000084
passing only known class data D L And training the classification network SKResNet to obtain a closed set identification network. Extracting feature F P Output activation vector of layer:
Figure BDA0003688143380000092
only the activation vectors which can be correctly classified in v are reserved and are divided according to different classes to form an activation vector set S k K is 1, 2. Each class is represented by the Mean Activation Vector (MAV):
μ k ＝mean(S k ) (15)
calculating S k Set of distances D (S) from the median activation vector to the Mean Activation Vector (MAV) k ,μ k ) In the present embodiment, the distance metric used is euclidean distance. Then D (S) k ,μ k ) The distance values in the sequence are sorted from small to large, the largest epsilon distance values of each type after sorting are selected, and according to an extreme value theory, the values of the epsilon distance values are subject to inverse Weibull distribution. Fitting the distance extreme value of each type of known sample to the corresponding MAV in the feature spaceObtaining K EVT models rho through the probability density function of the size k (θ k ,λ k ,κ k ). And in the identification stage, the distance between the identification sample and the MAV in the feature space is substituted into the EVT model, the reliability of the identification result is evaluated, and the reliability probability is given, wherein the farther the distance is, the lower the reliability is. And correcting the original neural network probability output through the reliability probability.
During the identification process, an open risk is introduced due to the addition of unknown samples. The open risk is defined as the ratio of the probability of identifying an unknown sample in open space as a known class to the probability of identifying a sample in full space as a known class:
Figure BDA0003688143380000091
o means open space, R means full space, f (x) is an identification function, f (x) > 0 when the sample is identified as a known class, and f (x) ═ 0 when the sample is identified as an unknown class. As can be seen from the definition of the open risk, the open risk is related to both the spatial range of the unknown class distribution and the discriminant function of the known class, so there are two ideas for the method of limiting the open risk: firstly, the size of O is limited, secondly, f (x) is restricted, the condition that the sample is judged to be a known class is strictly determined, and the probability of misjudgment in an open space is reduced. Theoretically, the open risk can always be limited to 0 by the method of setting the threshold value. However, in practical terms, not only the risk of openness but also the risk of experience due to classification errors are taken into account. Therefore, there is a need to minimize the risk of openness while controlling the risk of experience.
The OpenMax method uses an EVT theory to model the extreme value of the distance between the sample and the class center in the feature space, the identification function is closely related to the distribution of the sample in the feature space, and the farther the distance between the identification sample and the class center is, the lower the reliability of the identification result is. The introduction of centrloss during training can make the distribution of known classes in the feature space more compact, and unknown classes still appear randomly in the feature space. For the single-class sample classification problem, when the empirical risk is attempted to be controlled to be 0, all known classes need to be correctly identified, and the identification threshold set in this way can cause high confidence to the unknown class samples and introduce open risks. After the CenterLoss training is added, the OpenMax algorithm can mark a positive open space reduction in the feature space, and the conditions for identifying the samples in the open space as known classes are more severe, as shown in fig. 5. Thus, the introduced risk of openness is much reduced when the control experience risk is the same level when identifying.
A common classification network training uses cross entropy loss, and the loss function after improved training is composed of cross EntropyLoss and CenterLoss together. The cenerloss loss function is added during model training, so that the intra-class distance of the known class sample can be reduced, the range of the known class in the feature space is further limited, the open risk is reduced, and the method is suitable for improving the open set identification model adopted by the text. Equation (17) represents the loss of one training batch, N is the number of samples in one batch, F (x) i ) The feature vector extracted by the representation feature extraction network is the output feature vector of the first full connection layer;
Figure BDA0003688143380000101
and b i Representing the weight and offset of the second fully-connected layer;
Figure BDA0003688143380000102
representing the center of the class i sample in the feature space of the class to which they belong, as they vary during the training process. λ is a hyper-parameter that adjusts the crossEntropyLoss and CenterLoss weights. During training, since the centrloss takes the distance from the center of each batch sample as a loss, the intra-class distance of each class is continuously reduced.
Figure BDA0003688143380000103
Furthermore, in the embodiment of the scheme, the trained signal recognition model is used for recognizing radar radiation source individual signals to be recognized, firstly, the radar radiation source individual signals to be recognized are used as model input to obtain an activation vector of an input signal, and the distance between the activation vector and the mean value of the activation vector of each marked signal sample category is calculated; then, the mean value is brought into a corresponding extreme value theoretical model, a reliability score is obtained through the extreme value theoretical model, and the reliability score is used for adjusting the activation vector value; and finally, identifying the belonging radiation source category by normalizing various probabilities of the input signals.
The algorithm for radiation source identification can be designed as follows:
Figure BDA0003688143380000104
Figure BDA0003688143380000111
in the identification stage, a sample (x) to be identified is input into the trained network i ,y i )∈D R To obtain an activation vector v i . Separately calculate v i And a distance d of K known MAVs i,k . Then d is i,k Substitution into class k EVT model ρ k (θ k ,λ k ,κ k ) In the step (b), a reliability score ω is obtained i (k) In that respect For v i Are ordered from small to large, for ω i (k) Correcting the value of (c):
Figure BDA0003688143380000112
correction v i Is given a value of
Figure BDA0003688143380000113
Figure BDA0003688143380000114
Defining the components of an activation vector for newly added one-dimensional unknown classes
Figure BDA0003688143380000115
The value of (c):
Figure BDA0003688143380000116
recalculating the probability of each class after normalization:
Figure BDA0003688143380000117
wherein j-0 represents that the recognition result is an unknown class. And when the class corresponding to the maximum output probability value is an unknown class or the maximum probability value is smaller than a given threshold value delta, judging that the final recognition result is the unknown class. And in other cases, judging that the final identification result is the known class corresponding to the maximum probability value.
Further, based on the foregoing method, an embodiment of the present invention further provides a radar radiation source individual open-set identification system, including: a model training module and a signal recognition module, wherein,
the model training module is used for training the signal recognition model by constructing the signal recognition model and utilizing the marked signal samples in the radar signal database, wherein the transient sequence of the radar signal is used as the network input of the signal recognition model, the input signal characteristics are extracted through the signal recognition model network, and the extracted signal characteristics are classified and recognized;
and the signal identification module is used for inputting the radar radiation source individual signals to be identified intercepted by the receiver as a signal identification model, identifying the radar radiation source individual signals to be identified by using the trained signal identification model, if the radar radiation source individual signals are identified as a known class, outputting the name of the radiation source belonging to the known class, and if the radar radiation source individual signals are identified as an unknown class, marking the signals as the unknown class and adding the unknown class into a radar signal database so as to identify and mark the signals manually.
To verify the validity of the protocol, the following further explanation is made with reference to the test data:
7 different signal sources are adopted to simulate the signals transmitted by a radar transmitter, namely an AWG signal generator, an EXG signal source and 5 NI USRP-2901 with the same model. The signal source transmits a simple pulse signal with the carrier frequency of 800MHz and the pulse width of 1 us. Signals were acquired using a Tektronix high performance oscilloscope. The sampling rate is 50GHz, when 800MHz signals are collected, the range of instantaneous bandwidth is DC to 1GHz, and sampling rate is reduced by 10 times after collection. The signal source and the high-performance real-time oscilloscope are connected by the feeder line as a transmission channel, so that the acquired signal has a higher signal-to-noise ratio, and more characteristics caused by hardware difference among devices are reserved, but not the channel environment characteristics.
Adding noise with SNR of 26,30dB to the acquired signal, increasing the sample size, corresponds to data enhancement. Under each signal-to-noise ratio, randomly extracting 500 samples from each radiation source individual to form a training set; 100 processes are respectively extracted to form a verification set and a test set.
3 of the 7 acquired radiation source signals are selected as known classes to be trained to obtain a model, and 0,1,2 … and 5 of the other 5 radiation source signals are respectively extracted as unknown radiation sources. The test signal contains M K Class known class and M U Class unknown class, fixed M K K-3, i.e. containing all kinds of known classes, then different sources of the unknown class correspond to different openness ratios:
Figure BDA0003688143380000121
in the test experiment, all are
Figure BDA0003688143380000122
And (3) selecting 5 combinations of the known signal types to cover all the known signal types, wherein the combinations are { AWG, USRP1, USRP3}, { EXG, USRP2, USRP3}, { USRP1, USRP2, USRP3}, { USRP1, USRP4, USRP5}, { USRP2, USRP4 and USRP5 }. A recognition model can be obtained during the training phase using each combination of known class signals. In the testing stage, under the condition that each identification model corresponds to 5 openness degrees, the total openness degrees are calculated
Figure BDA0003688143380000123
A combination of test signals.
And in the model training process, signals with different signal-to-noise ratios are mixed and trained, and the test is carried out according to different signal-to-noise ratios in the test process. Under the condition of the same signal-to-noise ratio and the same openness ratio (namely the same number of unknown signals), the average value of the test results of different models and different unknown signal combinations is calculated, and then the openness ratio and the signal-to-noise ratio are traversed. The final experiment result, which is obtained by replacing 5 random seeds, monte carlo 5 times the above experiment process and calculating the mean value of the result, is shown in fig. 6. As can be seen, the recognition accuracy of the method is high, the ACC and the F1 values are maintained at a high level (greater than 0.9) under different openness degrees, and the method has good open set scene adaptability.
Unless specifically stated otherwise, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
Finally, it should be noted that: the above-mentioned embodiments are only specific embodiments of the present invention, which are used for illustrating the technical solutions of the present invention and not for limiting the same, and the protection scope of the present invention is not limited thereto, although the present invention is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive the technical solutions described in the foregoing embodiments or equivalent substitutes for some technical features within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the embodiments of the present invention, and they should be construed as being included therein. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A radar radiation source individual open set identification method is characterized by comprising the following contents:
constructing a signal recognition model, and training the signal recognition model by using marked signal samples in a radar signal database, wherein a transient sequence of a radar signal is used as signal recognition model network input, and the input signal characteristics are extracted and classified and recognized through the signal recognition model network;
the method comprises the steps of inputting radar radiation source individual signals to be identified, intercepted by a receiver, as signal identification models, identifying the radar radiation source individual signals to be identified by utilizing the trained signal identification models, if the radar radiation source individual signals are identified as known classes, outputting radiation source names of the known classes, and if the radar radiation source individual signals are identified as unknown classes, marking the signals as unknown classes, adding the unknown classes into a radar signal database, and identifying and marking manually.
2. The method for identifying individual open sets of radar radiation sources according to claim 1, wherein the signal identification model adopts a multi-scale ResNet network, and the multi-scale ResNet network comprises: the system comprises an SK module for extracting input multi-channel signal characteristics, a characteristic fusion module for performing fusion processing on the multi-channel signal characteristics, an average pooling layer for performing average pooling operation on the fusion characteristics to obtain characteristic vectors, a full connection layer for performing dimension processing on the characteristic vectors, a softmax layer for performing classification operation on the characteristic vectors after the dimension processing to obtain channel coefficients, and an output layer for fusing the channel coefficients and the corresponding channel signal characteristics extracted by the SK module and determining a final classification recognition result.
3. The radar radiation source individual open set identification method according to claim 2, characterized in that the SK module adopts a one-dimensional residual convolution structure, and the one-dimensional residual convolution structure includes a plurality of convolution kernels of different sizes and selection paths of a full connection layer, and the weights of the different selection paths are adjusted according to the loss in the training process.
4. The radar radiation source individual open set identification method according to claim 1, characterized in that in the training of the signal identification model, firstly, the signal identification model is trained by using marked signal samples to obtain a signal identification network of a closed set, and an output characteristic activation vector is obtained through a full connection layer; then, retaining the correctly classified activation vectors, and forming an activation vector set according to different classes; then, calculating a distance set from the activation vector to the activation vector mean under each category element in the activation vector set; and finally, arranging the distance values in the distance set in sequence, selecting the largest epsilon distance values of each type after sequencing, and obtaining an extreme value theoretical model corresponding to the type of the marked signal sample by utilizing an extreme value theory and fitting a probability density function of the size of the mean value distance extreme value of the activated vector of each type of the marked signal sample in the feature space.
5. The method for identifying the individual open set of radar radiation sources according to claim 1 or 4, wherein the target loss function trained by the signal identification model is represented as:
Figure FDA0003688143370000011
where N is the number of marked signal samples, F (x) i ) A feature vector representing a fully connected layer output;
Figure FDA0003688143370000012
and b i Representing full connection layer weights and offsets;
Figure FDA0003688143370000013
and the signal sample of the ith class is represented in the feature space center of the class to which the signal sample belongs, and lambda is a hyperparameter for adjusting the loss weight.
6. The method of claim 4, wherein Euclidean distance is used to measure the distance from the activation vector to the mean value of the activation vector.
7. The radar radiation source individual open set identification method according to claim 4, characterized in that, in identifying the radar radiation source individual signal to be identified by using the trained signal identification model, firstly, the radar radiation source individual signal to be identified is used as the model input to obtain the activation vector of the input signal, and the distance between the activation vector and the mean value of the activation vector of each labeled signal sample category is calculated; then, the mean value is brought into a corresponding extreme value theoretical model, a reliability score is obtained through the extreme value theoretical model, and the original activation vector value is adjusted by utilizing the reliability score; and finally, identifying the radiation source category by normalizing the probabilities of the various types of input signals.
8. The method for identifying individual open sets of radar radiation sources according to claim 7, wherein the reliability score is calculated by the following formula:
Figure FDA0003688143370000021
wherein K, s (j), v i Respectively representing the number of the known radiation source types, the serial number of the classification index of the activation vector with the size of j, the activation vector of the ith identification sample, and theta s(j) 、λ s(j) 、κ s(j) Three parameters of the s (j) th extremum model are respectively represented.
9. The method for identifying individual open sets of radar radiation sources according to claim 8, wherein the probability formula of each type of signal is expressed as:
Figure FDA0003688143370000022
wherein x is i Which is representative of the signal to be identified,
Figure FDA0003688143370000023
which is indicative of the class of the signal to be identified,
Figure FDA0003688143370000024
indicating the adjusted activation vector.
10. A radar radiation source individual open set identification system, comprising: a model training module and a signal recognition module, wherein,
the model training module is used for training the signal recognition model by constructing the signal recognition model and utilizing the marked signal samples in the radar signal database, wherein the transient sequence of the radar signal is used as the network input of the signal recognition model, the input signal characteristics are extracted through the signal recognition model network, and the extracted signal characteristics are classified and recognized;
and the signal identification module is used for inputting the radar radiation source individual signals to be identified intercepted by the receiver as a signal identification model, identifying the radar radiation source individual signals to be identified by using the trained signal identification model, if the radar radiation source individual signals are identified as a known class, outputting the name of the radiation source belonging to the known class, and if the radar radiation source individual signals are identified as an unknown class, marking the signals as the unknown class, adding the unknown class into a radar signal database, and identifying and marking the signals manually in the future.
CN202210652213.7A 2022-06-10 2022-06-10 Radar radiation source individual open set identification method and system Pending CN114970638A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210652213.7A CN114970638A (en) 2022-06-10 2022-06-10 Radar radiation source individual open set identification method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210652213.7A CN114970638A (en) 2022-06-10 2022-06-10 Radar radiation source individual open set identification method and system

Publications (1)

Publication Number Publication Date
CN114970638A true CN114970638A (en) 2022-08-30

Family

ID=82960953

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210652213.7A Pending CN114970638A (en) 2022-06-10 2022-06-10 Radar radiation source individual open set identification method and system

Country Status (1)

Country Link
CN (1) CN114970638A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116522242A (en) * 2023-04-28 2023-08-01 哈尔滨工程大学 Radiation source signal open set identification method based on diffusion model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116522242A (en) * 2023-04-28 2023-08-01 哈尔滨工程大学 Radiation source signal open set identification method based on diffusion model
CN116522242B (en) * 2023-04-28 2024-01-26 哈尔滨工程大学 Radiation source signal open set identification method based on diffusion model

Similar Documents

Publication Publication Date Title
CN113050042B (en) Radar signal modulation type identification method based on improved UNet3+ network
CN114564982B (en) Automatic identification method for radar signal modulation type
CN108596199A (en) Unbalanced data classification method based on EasyEnsemble algorithms and SMOTE algorithms
CN114692665B (en) Radiation source open set individual identification method based on metric learning
CN110501122A (en) A kind of adaptive leakage detection method of pressure vessel based on cluster
CN112749633B (en) Separate and reconstructed individual radiation source identification method
CN114372493B (en) Computer cable electromagnetic leakage characteristic analysis method
CN115114958A (en) Electromagnetic signal open set identification method based on supervised contrast learning
CN111310719B (en) Unknown radiation source individual identification and detection method
Dionelis et al. Tail of distribution GAN (TailGAN): GenerativeAdversarial-network-based boundary formation
CN112560596A (en) Radar interference category identification method and system
CN115294615A (en) Radio frequency fingerprint identification method based on machine learning
CN115546608A (en) Unmanned aerial vehicle data link electromagnetic interference classification and threat assessment method
CN116866129A (en) Wireless communication signal detection method
CN114970638A (en) Radar radiation source individual open set identification method and system
KR102407835B1 (en) Method and apparatus for classifying pulse radar signal properties based on machine learning
Tang et al. Specific emitter identification for IoT devices based on deep residual shrinkage networks
Han et al. Radar specific emitter identification based on open-selective kernel residual network
CN114487129A (en) Flexible material damage identification method based on acoustic emission technology
US20220124487A1 (en) Rf fingerprint signal processing device and rf fingerprint signal processing method
CN113378910A (en) Poisoning attack method based on electromagnetic signal modulation type identification of pure tag
CN114513374B (en) Network security threat identification method and system based on artificial intelligence
CN113534059B (en) Radar active interference identification method based on deep convolutional network under open set scene
CN115061094A (en) Radar target identification method based on neural network and SVM
Wang et al. FCM algorithm and index CS for the signal sorting of radiant points

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