Method and system for identifying individual open sets of radar emitters

Abstract

The present application belongs to the technical field of radar emitter identification, and particularly relates to a radar emitter individual open set identification method and system, a signal identification model is constructed, and the signal identification model is trained by using labeled signal samples in a radar signal database, wherein a transient sequence of the radar signal is used as network input of the signal identification model, signal features of the input are extracted by the signal identification model network, and the extracted signal features are classified and identified; the receiver intercepts a radar emitter individual signal to be identified as signal identification model input, the trained signal identification model is used to identify the radar emitter individual signal to be identified, if the identification is a known class, a name of the radar emitter of the known class is output, if the identification is an unknown class, the signal is labeled as the unknown class, and is added to the radar signal database for identification and labeling by manual operation. The present application has good generalization and fast identification speed, and has the abilities of classifying known emitters and detecting new unknown emitters, and is convenient for practical scene application.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar emitter identification, and particularly relates to a radar emitter individual open set identification method and system. BACKGROUND

[0002] The current electromagnetic environment presents a trend of increasing complexity, with increasing signal density and serious time and frequency domain aliasing. Radar plays an increasingly important role in life, especially with the accelerated popularization and application of automatic driving, radar imaging, radar detection and other technologies, radar signals have become an important part of the electromagnetic environment. In the face of complex electromagnetic environment, the traditional emitter identification method based on pulse description word has performance bottlenecks and cannot distinguish different individual radars of the same model. At the same time, due to the development of software radio technology, radar parameters are variable, making it more difficult to identify emitters through traditional parameters.

[0003] Under this background, many existing methods analyze radar signals to extract features that can reflect the hardware differences of different transmitters, achieving radar emitter individual identification. These features are also called fingerprint features, which are caused by the process gap of the hardware components of the emitter. Similar to human fingerprints, emitter fingerprint features are also difficult to modify unless the hardware changes. Therefore, emitter individual identification through fingerprint features is a good idea. Some researches extract features that can distinguish specific emitters according to expert knowledge, such as intra-pulse unintentional phase modulation features and unintentional amplitude modulation features, and then combine a classifier for classification. The advantages of this method are that the feature calculation has small computational complexity and fast speed; the disadvantages are that it requires strong expert knowledge background, the feature design process is cumbersome, and the generalization of the features is not strong. With the development of artificial intelligence technology, more and more researchers use deep learning to replace manual extraction of fingerprint features. Some methods need to perform complex preprocessing on the signal before inputting it into the neural network, which takes a long time to preprocess the signal during identification and has low identification efficiency. Some other methods directly use the original signal data as the network input, but they do not consider the characteristics of radar signal fingerprint features, resulting in a certain degree of input redundancy and causing a large amount of calculation for the subsequent feature extraction network. In addition, existing methods rarely consider the openness of the real electromagnetic environment. Especially for methods based on discriminant functions, machine learning and deep learning, if the identified signal comes from a new emitter without prior information, it may be misclassified as a known class. This is a big loophole for spectrum monitoring systems that need to detect illegal users. SUMMARY

[0004] To this end, the application provides a radar radiation source individual open set identification method and system, which has good generalization and fast identification speed, and has the ability to classify known radiation sources and detect new unknown radiation sources, facilitating practical scene application.

[0005] According to the design scheme provided by the application, a radar radiation source individual open set identification method is provided, which includes the following contents:

[0006] A signal identification model is constructed, and the signal identification model is trained by using the labeled signal samples in the radar signal database, wherein the transient sequence of the radar signal is used as the network input of the signal identification model, the signal features of the input are extracted through the signal identification model network, and the extracted signal features are classified and identified;

[0007] The receiver intercepts the radar radiation source individual signal to be identified as the signal identification model input, and uses the trained signal identification model to identify the radar radiation source individual signal to be identified, if it is identified as a known class, and outputs the name of the known class, if it is identified as an unknown class, the signal is marked as an unknown class, and added to the radar signal database, so as to be identified and marked by artificial.

[0008] As the radar radiation source individual open set identification method of the application, further, the signal identification model adopts a multi-scale ResNet network, which includes: an SK module for extracting input multi-channel signal features, a feature fusion module for fusing the multi-channel signal features, an average pooling layer for performing average pooling operation on the fused features to obtain a feature vector, a full connection layer for dimension processing of the feature vector, a softmax layer for performing classification operation on the dimension-processed feature vector to obtain a channel coefficient, and an output layer for fusing the channel coefficient and the corresponding channel signal feature extracted by the SK module and determining the final classification and identification result.

[0009] As the radar radiation source individual open set identification method of the application, 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 channels for extracting signal features through convolution operation.

[0010] As the radar emitter individual open set identification method of the application, further, in the signal identification model training, first, the signal identification model is trained by using the labeled signal samples, the signal identification network of the closed set is obtained, and the output feature activation vector is obtained through the full connection layer; then, the correctly classified activation vectors are retained, and the activation vector sets are formed according to different categories; then, the distance set of the activation vector to the activation vector mean under each category element in the activation vector set is calculated; finally, the distance values in the distance set are arranged in order, the maximum ε distance values of each category after sorting are selected, and the extreme value theory model of the corresponding labeled signal sample category is obtained by fitting the probability density function of the extreme value size of the activation vector mean of each labeled signal sample in the feature space.

[0011] As the radar emitter individual open set identification method of the application, further, the target loss function of the signal identification model training is represented as: Wherein, N is the number of labeled signal samples, F(x i ) represents the feature vector output by the full connection layer; and b i represent the full connection layer weight and bias; represents the feature space center of the i-th signal sample in the category, and λ is a hyperparameter for adjusting the loss weight.

[0012] As the radar emitter individual open set identification method of the application, further, the Euclidean distance is used to measure the distance of the activation vector to the activation vector mean.

[0013] As the radar emitter individual open set identification method of the application, further, the trained signal identification model is used to identify the radar emitter individual signal to be identified, first, the radar emitter individual signal to be identified is taken as the model input to obtain the activation vector of the input signal, and the distance between the activation vector and the activation vector mean of each labeled signal sample category is calculated; then, the mean is brought into the corresponding extreme value theory model, the reliability score is obtained through the extreme value theory model, and the activation vector value is adjusted by using the reliability score; finally, the belonging source category is identified by normalizing the probability of each category of the input signal.

[0014] As the radar emitter individual open set identification method of the application, further, the reliability score calculation formula is represented as: Wherein, K, s(j), v i respectively represent the number of known category emitters, the activation vector classification index sequence number of the activation vector with the size order of j, and the activation vector of the i-th identification sample, θ s(j) , λ s(j) , κ s(j) respectively represent three parameters of the s(j)-th extreme value model.

[0015] As the radar emitter individual open set identification method of the application, further, the signal probability formula of each type is expressed as: Wherein, x i represents the signal to be identified, represents the signal type to be identified, represents the adjusted activation vector value.

[0016] Further, the application also provides a radar emitter individual open set identification system, comprising: a model training module and a signal identification module, wherein,

[0017] The model training module is used for training the signal identification model by constructing the signal identification model and using the labeled signal samples in the radar signal database, wherein the transient sequence of the radar signal is used as the network input of the signal identification model, the signal features of the input are extracted through the signal identification model network, and the extracted signal features are classified and identified;

[0018] The signal identification module is used for inputting the radar emitter individual signal to be identified intercepted by the receiver into the signal identification model, identifying the radar emitter individual signal to be identified by using the trained signal identification model, outputting the name of the known class if the identification is the known class, and adding the signal to the radar signal database if the identification is the unknown class, so as to identify and mark by manual operation.

[0019] The application has the following beneficial effects:

[0020] The application can extract the emitter fingerprint features, classify the known emitter equipment, identify the unknown emitter at the same time, extract the emitter fingerprint features by using the one-dimensional convolutional neural network, does not depend on the expert knowledge, has good generalization performance, uses the SKResNet with the residual network structure, prevents the training degradation, and the different receptive fields of the multi-scale convolution kernel, can dynamically adjust the receptive field according to the input and the loss, can extract the features with good representation, can extract the transient sequence of the radar signal in the preprocessing process, the calculation process is efficient, compared with directly inputting the complete pulse signal into the neural network, more saves the calculation and storage resources, and considers that the real signal environment is open, improves the SKResNet network based on the OpenMax method, makes it have the open set identification ability, can identify the unknown emitter while correctly classifying the known class emitter, and adds the CenterLoss for training according to the principle of the OpenMax, and further improves the performance of the method. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is the radar emitter individual open set identification process schematic in the embodiment;

[0022] Figure 2 This is a schematic diagram of the signal preprocessing process in the embodiment;

[0023] Figure 3 This is a schematic diagram of the SK module structure in the embodiment;

[0024] Figure 4 This is a schematic diagram of the SKResNet structure in the embodiment;

[0025] Figure 5 This example illustrates the comparison of sample distribution in the feature space before and after training with CenterLoss.

[0026] Figure 6 The following is a schematic diagram of the test results of the solution in this embodiment. Detailed implementation method:

[0027] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0028] This invention provides a method for identifying individual open-set radar radiation sources, comprising the following:

[0029] A signal recognition model is constructed and trained using labeled signal samples from a radar signal database. The transient sequence of radar signals is used as the input to the signal recognition model network. The signal recognition model network is used to extract the input signal features and classify and recognize the extracted signal features.

[0030] The individual radar radiation source signals intercepted by the receiver are used as input to the signal recognition model. The trained signal recognition model is used to identify the individual radar radiation source signals. If the signal is identified as a known class, the name of the radiation source belonging to the known class is output. If the signal is identified as an unknown class, it is marked as an unknown class and added to the radar signal database for manual identification and marking.

[0031] See Figure 1 As shown, assuming there are K known similar radar radiating sources, where "similar" refers to radiating sources that cannot be identified using conventional PDW and intra-pulse modulation parameters, but can only be distinguished by individual source identification. In this scheme, during the model training phase, only the radar radiating source individual signals D from the already labeled database are used. L ={(x i ,y i ), i∈[1,N L ],y i Training is performed within the range [1,K], where y represents the radiation source number and i is the signal sample number. The training yields a recognition model. During the recognition phase, the receiver intercepts the individual radar radiation source signal D.R = {(x i ,y i ), i∈[1, N R ], y i ∈[1, M]}, total M classes, where M K classes belong to known radiation sources, and M U classes belong to unknown radiation sources. Through the trained model and the open set recognition algorithm, the intercepted signal is identified, if identified as a known class, the name of the known class radiation source to which the identification belongs is further given; if identified as a new unknown radiation source, it is marked as an unknown class, and added to the radar database to wait for subsequent manual marking by other methods, which has good generalization and simultaneously has the identification and detection capabilities of known and unknown radiation sources, facilitating the application in actual scenarios.

[0032] The purpose of signal preprocessing is to normalize the signal to adapt to the input requirements of the one-dimensional network. Common types of radar signals include pulse signals and continuous wave signals. In the present scheme, the focus is on pulse signals. The instantaneous process of pulse generation is considered to be the transient state of the radar signal. The transient state feature is affected by the structure of the transmission system and the hardware, and there are differences in the transient state features of different individual radiation sources, so it can be used as a fingerprint feature to distinguish individuals.

[0033] Since the input of the one-dimensional neural network is generally fixed in length, using a one-dimensional neural network to process the signal requires preprocessing the input signal to adapt to the input requirements. Common deep learning-based radiation source identification methods directly input the original signal data into the network, but the fingerprint features of time-domain radar signals are mainly concentrated in the unstable state when the rising edge changes, i.e., the rising edge transient part. If the complete pulse is used as input, it will produce a relatively large input redundancy, increasing the computing, storage burden, and increasing the model inference time. Therefore, in the present scheme, only the transient part of the radar signal is extracted as the network input, and the preprocessing process can be designed as follows:

[0034] Step 1: Orthogonal sampling of the intercepted radar signal to obtain a discrete signal; or performing Hilbert transform on the sampled radar signal:

[0035] z(n) = a(n) + jb(n), 1≤n≤N (1)

[0036] Step 2: Calculate the fourth moment of z(n):

[0037]

[0038] Step 3: Calculate the energy trajectory of the signal:

[0039]

[0040] Step 4: Max-min normalization of energy trajectory:

[0041]

[0042] Step 5: Thresholding q(n) to obtain N t pairs of indices of the start and end of transient signals, as shown in Figure 2

[0043]

[0044] Assuming the input length of the network is fixed as L, the transient sequence is obtained according to each pair of sequence indices:

[0045]

[0046] representing the floor function, the above formula indicates that the original data sequence z(n) is divided into transient sequences with a length of L, and each sequence is centered on the transient signal index coordinate and expanded forward and backward by the same length (or differs by 1 point).

[0047] Step 6: Normalization of the real and imaginary parts of s i for model training and recognition.

[0048] Further, in the embodiments of the present case, the signal recognition model adopts a multi-scale ResNet network, which includes: an SK module for extracting input multi-channel signal features, a feature fusion module for fusion processing of the multi-channel signal features, an average pooling layer for performing average pooling operation on the fused features to obtain a feature vector, a fully connected layer for dimension processing of the feature vector, a softmax layer for performing classification operation on the dimension-processed feature vector to obtain each channel coefficient, and an output layer for fusing each channel coefficient with the corresponding channel signal feature extracted by the SK module and determining the final classification recognition result. Further, the SK module adopts a one-dimensional residual convolution structure, and the one-dimensional residual convolution structure includes multiple different convolution kernel channels for extracting signal features through convolution operation.

[0049] ​The radar signal generation process is relatively complex and has non-stationary characteristics. The classic radar structure includes an oscillator, an amplifier, a filter, a transistor, etc. The hardware differences of different devices lead to unintentional modulation, i.e., fingerprint characteristics. These characteristics are very subtle, and therefore, the feature extraction network needs to have strong expression performance. Inspired by the working mechanism of the visual cortex, many methods hope that in the designed feature extraction network, neurons can have different receptive fields to improve the expression ability of the network. One-dimensional neural network performs one-dimensional convolution operation on the signal. One-dimensional convolution operation is similar to filtering, and different sizes of one-dimensional convolution kernels correspond to different sizes of filter coefficients, have signal analysis ability with different resolution, and are equivalent to learning signal features from different scale views. The residual network structure can effectively avoid the gradient explosion or disappearance problem of the neural network as the depth increases. When the recognition task is complex and the model complexity needs to be increased, the residual connection method has great advantages. Among them, the multi-scale ResNet network (Multi Scale Residual Network, MSResNet) can be applied to vibration signal fault detection. The influence of the features extracted by different scale convolution kernels on the final recognition result is different, and the weight of the feature of which resolution should be automatically adjusted according to the characteristics of the input signal. However, MSResNet directly splices different scale signal features, and the range of the receptive field is fixed. Therefore, inspired by the selective kernel network (Selective Kernel Network, SKNet), a multi-scale ResNet network with variable receptive field (Selective Kernel Residual Network, SKResNet) can be designed.

[0050] In the embodiment of the case, the SKResNet network model is mainly composed of an SK module, and the key of the SK module lies in that the network dynamically adjusts the weights of features of different convolution kernel channels. As shown in Figure 3 , the SK module can be designed to have three channels, and the module input features pass through 1x3, 1x5 and 1x7 one-dimensional residual convolution modules (ResBlock) in turn to obtain features with different receptive fields. When performing convolution operation, zero padding operation of appropriate size is performed to make the convolution operation not change the length of the feature vector. The ResBlock adopts the general residual network architecture, and its output is expressed as F i out :

[0051] F i out =F in +f(F in ,W i ),i=3,5,7 (7)

[0052] Fin W represents the input to ResBlock. i This represents the parameter value of the i-th ResBlock.

[0053] Then, the features from different branches are fused to obtain the fused features.

[0054]

[0055] For fusion feature F fused Average pooling is performed to obtain feature values ​​V0 across different channels. A fully connected network is used to reduce the dimension of F to C′, and then the dimension is increased to C, resulting in three feature vectors of the same size. This ensures that the feature selection path has good selection performance. A softmax operation is then performed on the three feature vectors along the channel dimension to generate the coefficient values ​​of the three feature paths in each channel.

[0056]

[0057] Compare the coefficient values ​​with the original feature F i out Multiply, then add and combine to obtain the selected feature Y. out :

[0058]

[0059] During training, the fully connected layers along the feature selection pathway adjust the weights of different branches based on the model loss. Without a feature selection pathway, it's equivalent to assigning the same weights to branches with different kernel sizes. The model mainly consists of three SKBlock layers and two fully connected layers (FC), as follows... Figure 4 As shown. The output of the first FC is a feature, and this fully connected layer is represented as F. P In this embodiment, F can be... P The output dimension is set to 2 to facilitate visualization of the feature distribution. The output dimension of the second FC is equal to the number K of known radiation sources.

[0060] Further, in the embodiment, in the signal recognition model training, firstly, the signal recognition model is trained by using the labeled signal samples, to obtain a closed set signal recognition network, and an output feature activation vector is obtained through a full connection layer; then, the correctly classified activation vectors are reserved, and the activation vector sets are formed according to different categories; then, the distance set of the activation vector to the activation vector mean in each category element in the activation vector set is calculated; finally, the distance values in the distance set are arranged in order, the maximum ε distance values of each category after the arrangement are selected, and the extreme value theory model of the corresponding labeled signal sample category is obtained by fitting the probability density function of the extreme value size of the activation vector mean of the labeled signal sample in the feature space of each category.

[0061] Extreme Value Theory (EVT) is often used to describe the probability of extreme events. A set of independent and identically distributed random variables x1, x2,..., x N satisfy the Fisher-Tippett theorem:

[0062] Let m = max{x1, x2,..., x N}, if there is a set of real numbers (a n , b n ) that satisfies any a n > 0 and and satisfies:

[0063]

[0064] When F(x) is a non-degenerate distribution function, F(x) must belong to the Gumbel family, Fréchet family or inverse Weibull family distribution. That is, the maximum value of a set of independent and identically distributed random variables converges to one of the above three distributions after proper normalization. The three distribution families are actually three cases of generalized extreme value distribution ξ>0, ξ=0, ξ<0, and the cumulative distribution function is:

[0065]

[0066] When ξ<0 in the above formula, it is an inverse Weibull distribution, and the cumulative probability distribution function is equivalent to:

[0067]

[0068] In the above formula, σ, ζ, θ>0.

[0069] The traditional closed set recognition network obtains the recognition probability of each class by normalizing the recognition score through the softmax layer. However, in the open set recognition process, the sum of the probabilities of the original to-be-classified classes is no longer 1 due to the influence of the unknown class. OpenMax improves the output of the softmax layer of the traditional closed set recognition network, increases the output dimension by one dimension, and gives the probability that the recognized class belongs to the unknown class. It uses the extreme value theory to model the distance between the feature vectors (also known as activation vectors) of the classification network. During recognition, the reliability of the original recognition score is re-evaluated and corrected according to the EVT distribution fitted during the recognition process, and finally open set recognition is performed.

[0070] In the embodiments of the present case, OpenMax is used to improve SKResNet to form Open-SKResNet, so that it has open set recognition capability. The training algorithm can be designed as follows:

[0071]

[0072] Only known class data D L Train the classification network SKResNet to obtain a closed set recognition network. Extract the feature F P The output activation vector of the layer is:

[0073]

[0074] Only the activation vectors that can be correctly classified in v are retained, and are divided according to different classes to form an activation vector set S k , k = 1, 2,..., K. Each class is represented by a mean activation vector (MAV):

[0075] μ k = mean (S k ) (15)

[0076] Calculate the distance set D(S k , μ k ) of the activation vectors in S k to the mean activation vector (MAV). In the embodiments of the present case, the distance metric used is the Euclidean distance. Then, the distance values in D(S k , μ k ) are sorted in ascending order, and the largest ε distance values of each class after sorting are selected. According to the extreme value theory, their values should obey the inverse Weibull distribution. Fit the probability density function of the extreme value size of the distance of each known sample in the feature space to its corresponding MAV, to obtain K EVT models ρ k (θ k , λ k , κ k). The recognition stage evaluates the reliability of the recognition result according to the distance between the recognition sample and the MAV in the feature space, substitutes the distance into the EVT model, and gives a reliability probability. The farther the distance, the lower the reliability. The original neural network probability output is corrected by the reliability probability.

[0077] In the recognition process, the open risk is introduced due to the addition of unknown samples. The open risk is defined as the ratio of the probability of recognizing the unknown sample in the open space as the known class to the probability of recognizing the sample in the whole space as the known class:

[0078]

[0079] O refers to the open space, R refers to the whole space, and f(x) is a recognition function. When the recognition sample is the known class, f(x)>0, and when the recognition sample is the unknown class, f(x)=0. According to the definition of the open risk, the open risk is related to the spatial range of the unknown class distribution and the known class discrimination function. Therefore, there are two ways to limit the open risk: one is to limit the size of O, and the other is to restrict f(x) and strictly determine the condition of recognizing the sample as the known class to reduce the probability of misjudgment in the open space. In theory, the threshold setting method can always limit the open risk to 0. However, in practical problems, not only the open risk needs to be considered, but also the experience risk brought by the classification error. Therefore, it is necessary to control the experience risk while trying to reduce the open risk.

[0080] The OpenMax method uses the EVT theory to model the extreme value of the distance between the sample and the class center in the feature space. The recognition function is closely related to the distribution of the sample in the feature space. The farther the distance between the recognition sample and the class center, the lower the reliability of the recognition result. The introduction of CenterLoss in the training process can make the distribution of the known class in the feature space more compact, and the unknown class is still randomly distributed in the feature space. For single-class sample classification problems, when trying to control the experience risk to 0, it is necessary to correctly recognize all known classes. The recognition threshold set in this way will lead to a high reliability for unknown class samples, introducing open risk. After adding CenterLoss training, the OpenMax algorithm reduces the open space that can be marked as positive in the feature space, and makes the condition of recognizing the sample in the open space as the known class more stringent, as shown in Figure 5 Therefore, when the experience risk is controlled at the same level during recognition, the introduced open risk is greatly reduced.

[0081] Common classification network training uses cross-entropy loss. The improved training loss function is composed of CrossEntropyLoss and CenterLoss. Adding the CenterLoss loss function during model training can reduce the intra-class distance of known class samples, further restricting the range of known classes in the feature space, which is beneficial to reducing open set risk and is suitable for improving the open set recognition model adopted in this paper. Equation (17) represents the loss of a training batch, where N is the number of samples in a batch, and F(x i ) represents the feature vector extracted by the feature extraction network, that is, the output feature vector of the first fully connected layer; and b i This represents the weights and biases of the second fully connected layer; Let represent the center of the i-th class sample in the feature space of its class, and let λ represent the changes during training. λ is a hyperparameter that adjusts the weights of CrossEntropyLoss and CenterLoss. During training, since CenterLoss uses the distance between the sample and the center within each batch as the loss, the intra-class distance of each class continuously decreases.

[0082]

[0083] Furthermore, in this embodiment, the trained signal recognition model is used to identify individual radar radiation source signals. First, the individual radar radiation source signals to be identified are used as model inputs to obtain the activation vectors of the input signals, and the distance between the activation vectors and the mean of the activation vectors of each labeled signal sample category is calculated. Then, the mean is substituted into the corresponding extreme value theory model to obtain a reliability score, and the reliability score is used to adjust the activation vector values. Finally, the radiation source category is identified by normalizing the probabilities of each type of input signal.

[0084] The algorithm for radiation source identification can be designed as follows:

[0085]

[0086]

[0087] During the recognition phase, the sample to be recognized (x) is input into the trained network. i ,y i )∈D R The activation vector v is obtained. i Calculate v separately. i The distance d between the MAV and the K known classes i,k Then d i,k Substituting into the k-th type of EVT model ρ k (θ k, lambda k , kappa k ), the reliability score omega i (k) is obtained. The values of omega i (k) are sorted in ascending order, and the values of v i are modified:

[0088]

[0089] The values of v i are modified, and

[0090]

[0091] The value of the activation vector component of the newly added one-dimensional unknown class is defined as:

[0092]

[0093] The probabilities of each class after normalization are recalculated:

[0094]

[0095] Where j=0 represents the recognition result as an unknown class. When the class corresponding to the maximum output probability is an unknown class, or the maximum probability is less than a given threshold delta, it is determined that the final recognition result is an unknown class. In other cases, it is determined that the final recognition result is the known class corresponding to the maximum probability.

[0096] Further, based on the above method, the embodiment of the application also provides a radar radiation source individual open set recognition system, comprising: a model training module and a signal recognition module, wherein,

[0097] The model training module is used to train the signal recognition model by constructing the signal recognition model and using the labeled 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 signal features of the input are extracted through the signal recognition model network, and the extracted signal features are classified and recognized;

[0098] The signal recognition module is used to input the radar radiation source individual signal to be recognized intercepted by the receiver into the signal recognition model, and identify the radar radiation source individual signal to be recognized by using the trained signal recognition model, if the radar radiation source individual signal to be recognized is identified as a known class, the name of the known class to which the radar radiation source individual signal to be recognized belongs is output, if the radar radiation source individual signal to be recognized is identified as an unknown class, the signal is marked as an unknown class, and added to the radar signal database, so as to be identified and marked by manual operation.

[0099] To verify the effectiveness of the scheme, the following test data are used for further explanation and description: ​

[0100] Seven different signal sources are used to simulate the radar transmitter, which are AWG signal generator, EXG signal source, and five NI USRP-2901s. The signal source transmits a simple pulse signal with a carrier frequency of 800 MHz and a pulse width of 1 us. A Tektronix high-performance oscilloscope is used to collect the signal. The sampling rate is 50 GHz, and when collecting a 800 MHz signal, the instantaneous bandwidth range is DC to 1 GHz. After collection, the sampling rate is reduced by 10 times. The signal source and high-performance real-time oscilloscope are connected by a feed line as a transmission channel, ensuring that the collected signal has a high signal-to-noise ratio and retains more features caused by hardware differences between devices rather than channel environment characteristics.

[0101] Noise with SNR = 26, 30 dB is added to the collected signal to increase the sample size, which is equivalent to data enhancement. Under each signal-to-noise ratio, 500 samples are randomly selected from each individual radiation source to form a training set. 100 samples are extracted to form a validation set and a test set, respectively.

[0102] From the seven collected radiation source signals, three of them are selected as known classes to train the model. From the remaining five, 0, 1, 2,..., 5 are extracted as unknown class radiation sources. The test signal contains M K known classes and M U unknown classes, with M K = K = 3, i.e., all known classes are included, then different unknown class radiation sources correspond to different openness rates:

[0103]

[0104] In the test experiment, there are known class signal combinations, and five of them are selected to cover all known class signals: {AWG, USRP1, USRP3}, {EXG, USRP2, USRP3}, {USRP1, USRP2, USRP3}, {USRP1, USRP4, USRP5}, {USRP2, USRP4, USRP5}. Using each known class signal combination, an identification model can be obtained in the training phase. In the test phase, each identification model corresponds to five openness degrees, a total of test signal combinations.

[0105] In model training, different signal-to-noise ratios are mixed together for training, and testing is performed according to different signal-to-noise ratios. Under the same signal-to-noise ratio, the same openness rate (i.e., the same number of unknown signals), different models, and different unknown class signal combinations are tested. The average value of the test results is calculated, and then the openness rate and signal-to-noise ratio are traversed. Replace the five random seeds, and perform the above experimental process five times by Monte Carlo. The average value of the results is calculated as the final experimental result, as shown inFigure 6 It can be seen that the recognition accuracy of the method is high, and the ACC and F1 values are maintained at a high level (greater than 0.9) under different open degrees, and the method has good open set scene adaptability.

[0106] Unless specifically stated, the relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0107] Finally, it should be noted that the above-described embodiments are only specific implementations of the present application, which are used to illustrate the technical solutions of the present application, but not to limit it. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for some technical features within the technical scope disclosed by the present application. The modification, change or substitution does not make the corresponding technical solution deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for identifying individual open sets of radar radiation sources, characterized in that, Comprise as follows: The signal recognition model is constructed and trained by using the labeled 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 signal features of the input are extracted by the signal recognition model network, and the extracted signal features are classified and recognized; the signal recognition model adopts a multi-scale ResNet network, which includes an SK module for extracting input multi-channel signal features, a feature fusion module for fusing the multi-channel signal features, an average pooling layer for performing an average pooling operation on the fused features to obtain a feature vector, a fully connected layer for dimension processing of the feature vector, a softmax layer for performing a classification operation on the dimension-processed feature vector to obtain channel coefficients, and an output layer for fusing the channel coefficients and the corresponding channel signal features extracted by the SK module to determine the final classification and recognition result; the SK module adopts a one-dimensional residual convolution structure, which includes multiple convolution kernels of different sizes and a selection channel with a fully connected layer, and the weights of different selection channels are adjusted according to the loss of the training process; in the training of the signal recognition model, firstly, the original radar data sequence is divided into transient sequences with a length of L, each sequence is centered on the transient signal index coordinate and expanded forward and backward by the same length, the real and imaginary parts of the transient sequence are normalized to obtain signal samples, the signal recognition model is trained by using the labeled signal samples, a closed-set signal recognition network is obtained, and the output feature activation vector is obtained through the fully connected layer; then, the correctly classified activation vectors are retained, and the activation vector set is formed according to different categories; next, the distance set of the activation vector to the mean value of the activation vector in each category element in the activation vector set is calculated; finally, the distance values in the distance set are arranged in order, the maximum ε distance values of each category after sorting are selected, and the extreme value theory is used to obtain the extreme value theory model of the corresponding labeled signal sample category by fitting the probability density function of the extreme value size of the mean value distance of the activation vector of each labeled signal sample in the feature space; the target loss function of the signal recognition model training is represented as: N is the number of labeled signal samples, F(x i ) represents the feature vector output by the fully connected layer; and b i represent the weights and bias of the fully connected layer; represents the center of the feature space of the i-th signal sample in the category, and λ is a hyperparameter for adjusting the loss weight; using the trained signal recognition model to identify the individual signal of the radar radiation source to be identified, first, the individual signal of the radar radiation source to be identified is taken as the model input to obtain the activation vector of the input signal, and the distance between the activation vector and the average of the activation vectors of each labeled signal sample in the category is calculated; then, the average is brought into the corresponding extreme value theory model, and the reliability score is obtained through the extreme value theory model, and the original activation vector value is adjusted by using the reliability score; finally, the category of the corresponding radiation source is identified by normalizing the probabilities of each category of the input signal. The receiver intercepts the radar radiation source individual signal to be identified as a signal recognition model input, uses the trained signal recognition model to identify the radar radiation source individual signal to be identified, if it is identified as a known class, and outputs the radiation source name of the known class, if it is identified as an unknown class, the signal is marked as an unknown class, and added to the radar signal database, so as to identify the mark by manual.

2. The radar emitter individual set identification method of claim 1, wherein, The Euclidean distance is used to measure the distance of the activation vector to the mean of the activation vector.

3. The method of claim 1, wherein: The reliability score calculation formula is represented as: Wherein, K, s(j), v i respectively represent the known class radiation source type number, the activation vector classification index serial number with the size order of j, and the activation vector of the i th identification sample, θ s(j) , λ s(j) , κ s(j) respectively represent three parameters of the s(j) th extreme value model.

4. The radar emitter individual set identification method of claim 3, wherein, The signal probability formula of each type is expressed as: wherein x i represents a signal to be identified, represents a signal type to be identified, represents an adjusted activation vector.

5. A radar emitter source individual open set identification system, characterized in that, The method is realized based on claim 1, comprising a model training module and a signal recognition module, wherein, The model training module is used to train the signal recognition model by constructing the signal recognition model and using the labeled 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 features are extracted by the signal recognition model network, and the extracted signal features are classified and recognized; The signal recognition module is used to input the receiver intercepts the radar radiation source individual signal to be identified as a signal recognition model input, uses the trained signal recognition model to identify the radar radiation source individual signal to be identified, if it is identified as a known class, and outputs the radiation source name of the known class, if it is identified as an unknown class, the signal is marked as an unknown class, and added to the radar signal database, so as to identify the mark by manual.

Images