An Open Set Recognition Method for Radar Signals Based on Adaptive Seabird Algorithm

Abstract

This application discloses a radar signal open-set recognition method based on an adaptive seabird algorithm, belonging to the field of signal recognition technology. It constructs a radar signal sample library by selecting reliable samples through a sample screening algorithm, and then builds an adaptive threshold based on this library to improve the recognition rate and enhance its universality. Simultaneously, an adaptive seabird algorithm is designed to optimize the adaptive threshold weights, constructing an exploration direction factor. Combining spiral optimization search and linear optimization search expands the search range and strengthens the search capability. Adaptive weights and adaptive step sizes that change with the number of iterations are added to balance global and local search capabilities, improving the reliability of the radar signal open-set recognition method. Furthermore, an LSTM network is designed and trained to extract features and obtain embedding vectors, enhancing the network's generalization and adaptability, thus improving the accuracy and universality of the radar signal open-set recognition method and effectively increasing the accuracy of unknown signal recognition.

Description

Technical Field

[0001] This application relates to a radar signal open set recognition method based on an adaptive seabird algorithm, belonging to the field of signal recognition technology. Background Technology

[0002] Traditional radar signal recognition methods based on deep neural networks are mostly closed-set recognition, which assumes that all training data is available and uses all data for training, resulting in a model suitable for the task. However, with the increasing number of radar signals with complex modulation types, traditional closed-set recognition methods cannot effectively and promptly identify new types of radar signals when they appear.

[0003] Metric learning classifies samples based on the distance between them in the feature embedding space, which enhances adaptability. However, the features of samples of the same class are not clustered enough, the features of samples of different classes are not dispersed enough, the boundaries between categories are blurred, and the determination of the threshold depends on expert experience, resulting in poor reliability and low recognition efficiency.

[0004] Zhang Shao, in his paper "Signal Identification of Unknown Modulation Types Based on Deep Learning," uses the idea of ​​metric learning to identify unknown signals. However, its discrimination requires the accumulation of unknown signal samples, and the recognition accuracy is low. Lin Wanjie, in his paper "Research on Individual Identification of Unknown Radiation Sources Based on Deep Learning," improves the classic Triplet Loss algorithm in deep metric learning to solve the unknown discrimination problem in individual radiation source identification. However, its decision threshold is selected based on experience when making distance decisions, resulting in insufficient reliability and lack of generalization. Summary of the Invention

[0005] The purpose of this application is to provide a radar signal open set recognition method based on the adaptive seabird algorithm, which solves the problem of low recognition rate and poor generalization caused by poor selection of empirical thresholds when recognizing radar with low probability of intercept. This method enhances the reliability, generalization and adaptability of the recognition method and effectively improves the accuracy of unknown radar signal recognition.

[0006] To achieve the above objectives, the first aspect of this application provides a radar signal open-set recognition method based on an adaptive seabird algorithm, comprising:

[0007] Obtain a training set, a validation set, and a test set, wherein the training set includes several known radar signals, and the validation set and the test set each include several unknown radar signals;

[0008] Construct an LSTM network, train the LSTM network based on the training set, and obtain a feature extraction network;

[0009] The embedding vectors of the training set and the validation set are obtained according to the feature extraction network, and the UMAP algorithm is used to reduce the dimensionality of the embedding vectors of the training set and the validation set to obtain the corresponding low-dimensional features.

[0010] Based on the low-dimensional features of the training set, a distance-based sample selection algorithm is used to select reliable samples to construct a known radar signal sample library.

[0011] Based on the low-dimensional features of the sample library and the validation set, an adaptive threshold for the identification of unknown radar signals is constructed.

[0012] The adaptive threshold is optimized by designing an adaptive seabird algorithm. The adaptive weight and adaptive step size are designed by combining spiral optimization search and linear optimization search, and an exploration direction factor is introduced to perform iterative calculation to obtain the optimal weight.

[0013] The test set is used to identify unknown radar signals based on the feature extraction network and an adaptive threshold based on the optimal weights.

[0014] In one implementation, training the LSTM network based on the training set to obtain a feature extraction network includes:

[0015] Wavelet scattering transforms are performed on the known radar signals in the training set and the unknown radar signals in the validation set, respectively, to obtain the scattering characteristics of the training set and the scattering characteristics of the validation set.

[0016] The scattering features of the training set are used as training data to train the LSTM network to obtain the feature extraction network, wherein the feature extraction network includes an input layer, an LSTM layer, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0017] In one implementation, obtaining the embedding vectors of the training set and the validation set respectively based on the feature extraction network includes:

[0018] Perform the following steps on the scattering features of the training set and the scattering features of the validation set, respectively:

[0019] The corresponding scattering features are input to the input layer for initialization to obtain the output of layer 0. The output of layer 0 is input to the LSTM layer for calculation to obtain the output of layer 1. The output of layer 1 is input to the first fully connected layer for calculation to obtain the output of layer 2. The output of layer 2 is input to the second fully connected layer for calculation to obtain the output of layer 3. The output of layer 3 is used as the corresponding embedding vector.

[0020] In one implementation, the dimensionality reduction processing of the embedding vectors of the training set and the validation set using the UMAP algorithm includes:

[0021] Perform the following steps on the embedding vectors of the training set and the embedding vectors of the validation set, respectively:

[0022] Based on the embedding vector, the nearest neighbor data is obtained using the nearest neighbor algorithm;

[0023] Based on the neighboring data, the joint probability and its low-dimensional spatial distribution are calculated, and the cross-entropy of the joint probability and its low-dimensional spatial distribution is calculated as the loss function.

[0024] Based on the loss function, a dimensionality reduction model is obtained using the gradient descent algorithm;

[0025] According to the dimensionality reduction model, the embedding vector is mapped to a low-dimensional feature space to obtain the corresponding low-dimensional features.

[0026] In one implementation, the step of using a distance-based sample selection algorithm to select reliable samples to construct a known radar signal sample library includes:

[0027] The low-dimensional features of the training set are divided into k sets according to the labels, where k is the number of signal types;

[0028] Calculate the inter-class distances of each set to obtain the distance matrix;

[0029] Based on the distance matrix, reliable samples of each signal type are selected using a distance-based sample selection algorithm to form the known radar signal sample library.

[0030] In one implementation, constructing the adaptive threshold for unknown radar signal identification includes:

[0031] Calculate the class centers of the reliable samples;

[0032] Calculate the Euclidean distance from the low-dimensional features of the validation set to the class center;

[0033] The adaptive threshold is constructed based on the Euclidean distance.

[0034] In one implementation, the step of optimizing the adaptive threshold by designing an adaptive seabird algorithm, combining spiral optimization search and linear optimization search, and designing adaptive weights and adaptive step sizes, while introducing an exploration direction factor for iterative calculation to obtain the optimal weights includes:

[0035] An initial population containing several courtship pavilions is randomly generated. The following steps are iteratively performed on the initial population until the globally optimal solution is obtained as the optimal weight:

[0036] Calculate the initial position of each courtship pavilion and the attraction and fitness values ​​of each pavilion;

[0037] By combining spiral optimization search and linear optimization search, the potential locations of each courtship pavilion are calculated;

[0038] Adaptive step size and adaptive weight are designed separately, and an exploration direction factor to maintain population diversity is added to obtain a new fitness value corresponding to the potential position.

[0039] The fitness value is compared with the new fitness value, and the position of each courtship pavilion is updated. At the same time, a dimension is randomly selected for position mutation update of some courtship pavilions in each courtship pavilion.

[0040] The mating pavilions obtained by position update and the mating pavilions obtained by position mutation update are combined to obtain a combined population. The cost function values ​​of all mating pavilions in the combined population are sorted from smallest to largest. The mating pavilions whose cost function values ​​meet the preset conditions are retained, and the other mating pavilions are eliminated.

[0041] In one implementation, the identification of unknown radar signals on the test set based on the feature extraction network and an adaptive threshold based on the optimal weights includes:

[0042] The embedding vectors of the test set are obtained based on the feature extraction network.

[0043] The UMAP algorithm is used to reduce the dimensionality of the embedding vectors of the test set and obtain the low-dimensional features of the test set.

[0044] Calculate the Euclidean distance from the low-dimensional features of the test set to the class center;

[0045] The test set is used to identify unknown radar signals based on the Euclidean distance and the adaptive threshold of the optimal weight.

[0046] A second aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the first aspect or any embodiment of the first aspect.

[0047] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the first aspect or any embodiment of the first aspect.

[0048] As can be seen from the above, this application provides a radar signal open-set recognition method based on the adaptive seabird algorithm, which solves the problem of low recognition rate and poor generalization caused by poor selection of empirical thresholds in low intercept probability radar recognition. A reliable sample library is constructed by selecting reliable samples through a sample screening algorithm. An adaptive threshold is then constructed based on the radar signal sample library to improve the recognition rate and enhance the universality of the radar signal open-set recognition method. Simultaneously, an adaptive seabird algorithm is designed to optimize the adaptive threshold weights, construct an exploration direction factor, and combine spiral optimization search and linear optimization search to expand the search range and enhance the search capability. Adaptive weights and adaptive step sizes that change with the number of iterations are added to balance global and local search capabilities, improving the reliability of the radar signal open-set recognition method. Furthermore, an LSTM network is designed and trained to extract features and obtain embedding vectors, enhancing the network's generalization and adaptability, improving the accuracy and universality of the radar signal open-set recognition method, and effectively increasing the accuracy of unknown signal recognition. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A schematic flowchart illustrating a radar signal open set identification method provided in an embodiment of this application;

[0051] Figure 2 This is a schematic diagram of an adaptive seabird algorithm provided in an embodiment of this application. Detailed Implementation

[0052] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0053] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0054] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0057] Example 1

[0058] This application provides a radar signal open-set recognition method based on an adaptive seabird algorithm, such as... Figure 1 As shown, the method includes:

[0059] S11: Obtain a training set, a validation set, and a test set, wherein the training set includes several known radar signals, and the validation set and the test set each include several unknown radar signals;

[0060] Optionally, known radar signals refer to radar signals of a known type, and unknown radar signals refer to radar signals of an unknown type. First, known radar signals are generated as samples to form the training set (train). At the same time, radar signals containing unknown radar signals are obtained as samples to form the validation set (validation) and the test set (test).

[0061] train={sig_train1(t),sig_train2(t),…,sig_train n (t),…,sig_train N (t)}

[0062] validation={sig_v1(t),sig_v2(t),…,sig_v n (t),…,sig_v N (t)}

[0063] test={sig_t1(t),sig_t2(t),…,sig_t n (t),…,sig_t N (t)}

[0064] In the formula, sig_train n (t) represents the nth signal in the training set, sig_v n (t) represents the nth signal in the validation set, sig_t n (t) represents the nth signal in the test set, where n = 1, 2, ..., N, N is the number of radar signals, and t represents the time domain.

[0065] S12: Construct an LSTM network, train the LSTM network based on the training set, and obtain a feature extraction network;

[0066] Optionally, training the LSTM network based on the training set to obtain the feature extraction network includes:

[0067] Wavelet scattering transform is performed on the known radar signals in the training set (train) and the unknown radar signals in the validation set (validation) to obtain the scattering features tf_tr of the training set and the scattering features tf_v of the validation set.

[0068] The scattering features tf_tr of the training set are used as training data to train the LSTM network. Simultaneously, the center loss and diversity loss can be combined for training to obtain the feature extraction network, which includes an input layer, an LSTM layer, a first fully connected layer, a second fully connected layer, and a softmax layer.

[0069] Furthermore, wavelet scattering transform is performed on the unknown radar signals in the above test set in the same manner to obtain the scattering features tf_t of the test set, where the sets of the above scattering features are as follows:

[0070] tf_tr={tf_tr1,tf_tr2,…,tf_tr i ,…,tf_tr I}

[0071] tf_v = {tf_v1, tf_v2, ..., tf_v} i ,…,tf_v I}

[0072] tf_t = {tf_t1, tf_t2, ..., tf_t} i ,…,tf_t I}

[0073] In the formula, tf_tr i Let tf_v be the i-th training sample in the set tf_tr. i Let tf_t be the i-th validation sample in set tf_v. i Let tf_t be the i-th test sample in the set tf_t, where i = 1, 2, ..., I, and I is the number of samples.

[0074] In one implementation, the corresponding scattering features are obtained through wavelet scattering transform, which reduces the size of the input data, reduces intra-class differences, and preserves the distinguishability between classes. The process is as follows:

[0075] Calculate the scattering output S of the first layer J [φ]x:

[0076] S J [φ]x=x*φ J

[0077] In the formula, "*" represents the convolution operation, x is the input signal, and φ J It is a scale of 2 J A Gaussian low-pass filter.

[0078] Calculate the wavelet mode coefficients U of the first layer J [λ1]x:

[0079] U J [λ1]x=|x*ψ λ1 |

[0080] In the formula, U J For wavelet modulus operators, "|·|" represents the modulus operation, ψ λ1 Let λ denote the complex wavelet operator, and λ1 denote the first layer scattering propagation path.

[0081] Calculate the scattering output S of the second layer J [λ1]x:

[0082] S J [λ1]x=|x*ψ λ1 |*φ J

[0083] In the formula, S J It is a scattering operator.

[0084] Training set signal sig_train i (t) After multi-layer wavelet scattering, the wavelet scattering feature tf_tr is obtained. i Its composition is:

[0085] {S J [φ]x,S J [λ1]x,…,S J[λ np ]x}

[0086] In the formula, λ np S represents the scattering propagation path of the np-th layer. J It is a scattering operator.

[0087] S13: Obtain the embedding vectors of the training set and the validation set respectively according to the feature extraction network, and use the UMAP algorithm to perform dimensionality reduction processing on the embedding vectors of the training set and the validation set respectively to obtain the corresponding low-dimensional features;

[0088] Optionally, obtaining the embedding vectors of the training set and the validation set respectively based on the feature extraction network includes:

[0089] S131: Perform the following steps on the scattering features of the training set and the scattering features of the validation set respectively:

[0090] The corresponding scattering features are input to the input layer for initialization to obtain the output of layer 0. The output of layer 0 is input to the LSTM layer for calculation to obtain the output of layer 1. The output of layer 1 is input to the first fully connected layer for calculation to obtain the output of layer 2. The output of layer 2 is input to the second fully connected layer for calculation to obtain the output of layer 3. The output of layer 3 is used as the corresponding embedding vector.

[0091] In one implementation, the LSTM network is specifically an LSTM-DN network. The trained network model includes an input layer, an LSTM layer, two fully connected layers, and a softmax layer. The LSTM layer contains 150 hidden units, which mainly consist of three parts: a forget gate f... ti Input gate i ti and output gate o ti The specific process of obtaining the embedding vector of the training set based on the feature extraction network is as follows:

[0092] S1311: Initialize the weight matrix u of the network input layer 0 and bias coefficient matrix q 0 To obtain the output of layer 0 tf_tr i Let i be the i-th training sample.

[0093] S1312: Use the output of step S1311 as the input of the LSTM hidden unit to calculate the forget gate f. ti The output of is represented as follows:

[0094] f ti =σ(W f ·[C ti-1 ,hti-1 ,x ti ]+b f )

[0095] In the formula, x ti To hide the input of the unit at the current time ti, C ti-1 h represents the state of the previous unit. ti-1 W is the output of the previous unit. f Let b be the forget gate weight matrix. f Let σ be the forget gate bias matrix, and σ be the sigmoid activation function.

[0096] S1313: Combined with the output h of the previous unit ti-1 The input x at the current moment ti and the previous unit state C ti-1 Calculate the temporary state of the current cell. and input gate i ti Output:

[0097]

[0098] i ti =σ(W ti ·[C ti-1 ,h ti-1 ,x ti ]+b ti )

[0099] In the formula, W C Let b be the weight matrix of the memory cells. C W is the bias matrix for memory cells. ti Let b be the input gate weight matrix. ti This is the input gate bias matrix.

[0100] S1314: Calculate the current element state C by combining the current element's temporary state and the previous element's state. ti :

[0101]

[0102] In the formula, C ti-1 This refers to the memory state of the previous hidden unit. f represents the temporary memory state of the current unit. ti For the forget gate of the current unit, i ti This is the output gate of the current cell.

[0103] Combined with the current cell state C ti The output h of the previous unit ti-1 With the current input x ti Calculate the output gate o ti With the current unit output hti :

[0104] o ti =σ(W o ·[C ti ,h ti-1 ,x ti ]+b o )

[0105] h ti =o ti ·tanh(C ti )

[0106] In the formula, W o Let b be the output gate weight matrix. o This is the output gate bias matrix.

[0107] After calculating the output of multiple hidden units, the first layer output of the i-th training sample is obtained.

[0108] S1315: Use the output of S1314 as the input of the first fully connected layer to obtain the output.

[0109]

[0110] In the formula, For the i-th training sample, the output of the first layer at the index position is u. 2,index and q 2.index These are the weights and biases for the second layer and the index-th input, respectively, and N1 is the output. The length.

[0111] S1316: Use the output of S1315 as the input of the second fully connected layer to obtain the output e. i e i The output vector of the i-th training sample data in layer 3 is used as the embedding vector extracted by the network:

[0112]

[0113] In the formula, For the output of the second layer and the index 1 position of the i-th training sample, u 3,index1 and q 3,index1 These represent the weights and biases of the input at the 3rd layer and the 1st index, respectively, and N2 represents the output. The length.

[0114] S1317: e i The input is fed into the softmax layer, and the output of the softmax layer is calculated.

[0115]

[0116] In the formula, e i This is the output vector of the i-th training sample data in the 3rd layer. Let v be the classification probability of the i-th training sample, v represent the predicted class, v∈(1,k), k represent the number of radar signal types, and softmax(·) is the softmax activation function.

[0117] Optionally, the dimensionality reduction processing of the embedding vectors of the training set and the validation set using the UMAP algorithm includes:

[0118] S132: Perform the following steps on the embedding vectors of the training set and the embedding vectors of the validation set respectively:

[0119] Based on the embedding vector, the nearest neighbor data is obtained using the nearest neighbor algorithm;

[0120] Based on the neighboring data, the joint probability and its low-dimensional spatial distribution are calculated, and the cross-entropy of the joint probability and its low-dimensional spatial distribution is calculated as the loss function.

[0121] Based on the loss function, a dimensionality reduction model is obtained using the gradient descent algorithm;

[0122] According to the dimensionality reduction model, the embedding vector is mapped to a low-dimensional feature space to obtain the corresponding low-dimensional features.

[0123] In one implementation, the UMAP algorithm is used to reduce the dimensionality of the embedding vectors, resulting in smaller intra-class differences and larger inter-class distances in the lower-dimensional space. The process is as follows:

[0124] S1321: Use the embedding vectors in S131 as the input set: E = {e1, e2, ..., e...} i ,…,e I}, use the nearest neighbor algorithm to obtain each e i The set of m neighboring points Indicates with e i The most recent m-th point.

[0125] S1322: Calculate the nearest neighbor data ρ using the m-nearest neighbor set. i and σ i :

[0126]

[0127]

[0128] In the formula, ei and The distance is given by: min(·) is the minimum function, max(·) is the maximum function, exp(·) and log2(·) are the exponential and logarithmic functions respectively, and m is the number of neighboring points. Let j be the j-th neighboring point.

[0129] S1323: ρ obtained from S1322 i and σ i Calculate the computable association probability p j / i p j / i Represented as:

[0130]

[0131] Where p j / i The calculation is performed only on the m nearest neighbors, and on all other points p. j / i =0.

[0132] S1324: The correlation probability p from S1323 j / i The joint probability p can be obtained. ij :

[0133] p ij =(p j / i +p i / j )-p j / i p i / j

[0134] In the formula, p j / i For e i Choose e j p is the probability of a neighboring point. i / j For e j Choose e i The probability of being a neighboring point.

[0135] S1325: When mapping data to a low-dimensional space, the joint probability in S1324 will also be represented by the distribution q in the low-dimensional space. ij It can be represented as:

[0136]

[0137] In this case, a1 = 1.929 and a2 = 0.7915 are set.

[0138] S1326: Calculate p in S1324 and S1325 ij and q ij The cross-entropy, as the loss function L UMAP , can be represented as:

[0139]

[0140] S1327: Based on the loss function in S1326, the optimal dimensionality reduction model is obtained using the gradient descent algorithm, mapping the embedding vectors to the low-dimensional feature space Y = {y1,…,y}. i1 ,…,y I1}, where i1=1,2,…,I1, y i1 Let i be the i1th low-dimensional feature.

[0141] S14: Based on the low-dimensional features of the training set, reliable samples are selected using a distance-based sample selection algorithm to construct a known radar signal sample library.

[0142] Optionally, a distance-based sample selection algorithm is used to select reliable samples of type num to construct a known radar signal sample library, specifically including:

[0143] The low-dimensional features of the training set are divided into k sets according to the labels, and denoted as Y = {Y1, Y2, ..., Y...} c ,…,Y k}, c = 1, 2, ..., k, where k is the number of signal types;

[0144] Calculate the inter-class distances of each set to obtain the distance matrix;

[0145] Based on the distance matrix, reliable samples of each signal type are selected using a distance-based sample selection algorithm to form the known radar signal sample library.

[0146] In one implementation, taking any one of these sets as an example, the process of constructing a known radar signal sample library is represented as follows:

[0147] S141: Y c Given the set of feature vectors belonging to the c-th type of radar signal, find Y. c Normalized vector center

[0148]

[0149] In the formula, y b For set Y c The b-th sample, b = 1, 2, ..., m c m c For the set Y c The number of samples, ||·||2 is the L2 norm operation.

[0150] S142: Sets Y obtained from S141 c Calculate the inter-class distances of each set to obtain the distance matrix D. c , represents

[0151]

[0152] In the formula, For y b With y l The Euclidean distance between them, Y c Let m be the set of feature vectors belonging to the c-th type of radar signal. c For the set Y c The number of samples, and d b D c The row vector represented by the b-th row.

[0153] S143: Calculate D obtained from S142 c The smallest 2num values ​​in set Y, excluding the elements on the main diagonal, where num is the smallest value in the set Y. c Given the number of reliable samples, calculate their average value as d. mean Store the row and column indices into the index array G.

[0154] S144: Find the element that appears most frequently in G obtained from S143, and denote it as g. c The corresponding sample is denoted as y. c Calculate y c and distance d cs and with d mean Compare.

[0155] S145: If d in S144 cs <d mean Select d c The samples corresponding to the num-1 minimum values ​​in the set are taken as reliable samples for that class. Otherwise, the set Y is calculated. c Each sample to The distance is calculated, and the num-1 closest points are selected as reliable samples for that class.

[0156] S15: Based on the low-dimensional features of the sample library and the validation set, construct an adaptive threshold for the identification of unknown radar signals;

[0157] Optionally, the construction of the adaptive threshold for unknown radar signal identification includes:

[0158] Calculate the class centers of the reliable samples;

[0159] Calculate the Euclidean distance from the low-dimensional features of the validation set to the class center;

[0160] The adaptive threshold is constructed based on the Euclidean distance.

[0161] In one implementation, this application embodiment constructs an adaptive threshold based on a sample library, which improves the recognition rate of the algorithm and enhances its universality. The specific process of constructing the adaptive threshold is as follows:

[0162] S151: Calculate the class centers of various reliable samples in S145.

[0163]

[0164] In the formula, y id For set Y c The id-th reliable sample in the set Y, where id = 1, 2, ..., num, and num is the set Y. c The number of reliable samples, ||·||2 is the L2 norm operation.

[0165] S152: Obtain the embedding vectors from the validation samples in set tf_v in S12 through the steps described in S131, and map them to a low-dimensional sample space using the steps described in S132 to obtain the low-dimensional vector distribution Y. validation :

[0166] Y validation ={y1,…,y iv ,…,y Iv}

[0167] In the formula, y iv Let be the low-dimensional feature vector of the ivth validation sample.

[0168] S153: Calculate vector y iv To reliable sample class center Euclidean distance iv,c :

[0169]

[0170] S154: Constructing an adaptive threshold

[0171]

[0172]

[0173] In the formula, Let y be a vector iv The corresponding threshold, α c Distance is the weight for distance. iv,c Let y be a vector iv To the Reliability Center The distance is c, where c is the corresponding signal category.

[0174] S155: Samples that satisfy the following formula are classified as unknown signals:

[0175]

[0176] In the formula, Let y be the ivth sample vector. iv With any known class vector center The minimum value of the Euclidean distance.

[0177] S156: Based on the discriminant in S155, evaluate the validation set data to obtain the accuracy:

[0178]

[0179] Where TP represents the number of correctly identified cases and FP represents the number of incorrectly identified cases.

[0180] S16: The adaptive threshold is optimized by designing an adaptive seabird algorithm. The adaptive weight and adaptive step size are designed by combining spiral optimization search and linear optimization search. At the same time, the exploration direction factor is introduced to perform iterative calculation to obtain the optimal weight.

[0181] Optionally, an adaptive seabird algorithm is designed to optimize the adaptive threshold by weights. This combines spiral optimization search and linear optimization search to expand the search range and enhance search capabilities. Adaptive weights and adaptive step sizes are designed, which change with the number of iterations, balancing global and local search capabilities. Simultaneously, an exploration direction factor is introduced for iterative calculation to obtain the optimal weights. Specifically, the following steps are included:

[0182] An initial population containing several courtship pavilions is randomly generated. The following steps are iteratively performed on the initial population until the globally optimal solution is obtained as the optimal weight:

[0183] Calculate the initial position of each courtship pavilion and the attraction and fitness values ​​of each pavilion;

[0184] By combining spiral optimization search and linear optimization search, the potential locations of each courtship pavilion are calculated;

[0185] Adaptive step size and adaptive weight are designed separately, and an exploration direction factor to maintain population diversity is added to obtain a new fitness value corresponding to the potential position.

[0186] The fitness value is compared with the new fitness value, and the position of each courtship pavilion is updated. At the same time, a dimension is randomly selected for position mutation update of some courtship pavilions in each courtship pavilion.

[0187] The mating pavilions obtained by position update and the mating pavilions obtained by position mutation update are combined to obtain a combined population. The cost function values ​​of all mating pavilions in the combined population are sorted from smallest to largest. The mating pavilions whose cost function values ​​meet the preset conditions are retained, and the other mating pavilions are eliminated.

[0188] In one implementation, such as Figure 2 As shown, S16 specifically includes the following steps:

[0189] S161: Randomly initialize the population. Randomly generate an initial population containing NB courtship pavilions within the feasible region. The position of each pavilion is defined as DB dimension. The current iteration count is n_t, and the maximum iteration count is Maxlt. Initialize the position of the ni-th courtship pavilion:

[0190]

[0191] In the formula, Let U(0,1) represent the position of the ni-th mating pavilion in the nj-th dimension during the n_t-th iteration, where U(0,1) is a uniformly distributed random number within (0,1), and Smax is the position of the ni-th mating pavilion in the nj-th dimension. ni,nj and Smin ni,nj Let be the maximum and minimum values ​​of the ni-th mating pavilion in the nj-th dimension search interval, where ni = 1, 2, ..., NB, and nj = 1, 2, ..., DB.

[0192] S162: Calculate the attractiveness of each courtship pavilion. The attractiveness P of each courtship pavilion... i and fitness value ni Calculated using the following formula:

[0193]

[0194]

[0195] In the formula, fit ni Let f(s) represent the fitness values ​​of ni courtship pavilions. ni ) is the objective function of the ni-th mating pavilion.

[0196] S163: Combine spiral optimization search and linear optimization search to calculate the potential locations of the courtship pavilion. Its formula can be expressed as follows:

[0197]

[0198] in, Let s be the nj-th dimension component of the ni-th individual in the n_t-th generation; rd,nj Determined through a roulette wheel selection mechanism; elite,nj λ is the nj-th dimension component of the current optimal position of the entire population;nj It is the step size factor, w ni For inertial weights, con is a constant, μ1 is a random number in [-1,1], cos is a cosine function, and μ2 is a random number in [0,1].

[0199] S164: Optimize the line search method and design an adaptive step size λ nj :

[0200]

[0201] Where β is the maximum threshold of the step size; P nj This represents the probability of the target mating station being selected, ranging from 0 to 1. The step size is largest when the probability of the target position being selected is 0, and smallest when the probability of the target position being selected is 1. The step size decreases as the number of iterations increases, making it easier to approach the optimal solution.

[0202] S165: Optimize the spiral search method and design adaptive weights. wni :

[0203]

[0204] In the formula, μ3 is the variation factor, and the inertia weight w ni The overall trend is increasing, balancing the algorithm's global and local search capabilities.

[0205] S166: Add directional perturbation factor T r To maintain population diversity, it is represented as:

[0206]

[0207] T r = P(μ4) × sign(μ4)

[0208]

[0209]

[0210] In the formula, P x Given the stimulus probability, μ4 ​​is a uniformly distributed random number in [0,1], and T r The value of only fluctuates between 1, 0, and -1.

[0211] S167: Obtain the updated fitness value new_fit based on the formula in S162. ni , with fit ni The location of the courtship pavilion is updated by comparing the locations.

[0212]

[0213] S168: Randomly select one dimension for position mutation and update of some courtship pavilions. During this process, s ni,nj It follows a normal distribution, as shown below:

[0214]

[0215]

[0216] δ=z(Smax ni,nj -Smin ni,nj )

[0217] In the formula, δ is the standard deviation, z is the scaling factor, and Smax is the standard deviation. ni,nj and Smin ni,nj These are variables s ni,nj The upper and lower limits.

[0218] S169: Combined Population. At the end of each iteration, the old population and the population obtained from mutation are combined to form a combined population. All individuals in the combined population are sorted by cost function values ​​from smallest to largest, and the individual with the smallest cost function value is retained, while the others are eliminated. This process continues until the maximum number of iterations, MaxIt, is reached, yielding the global optimal solution and the optimal weights {α1,…,α}. c ,…,α k}

[0219] S17: Identify unknown radar signals on the test set based on the feature extraction network and the adaptive threshold based on the optimal weights.

[0220] Optionally, the step of identifying unknown radar signals in the test set based on the feature extraction network and an adaptive threshold based on the optimal weights includes:

[0221] The embedding vectors of the test set are obtained based on the feature extraction network.

[0222] The UMAP algorithm is used to reduce the dimensionality of the embedding vectors of the test set and obtain the low-dimensional features of the test set.

[0223] Calculate the Euclidean distance from the low-dimensional features of the test set to the class center;

[0224] The test set is used to identify unknown radar signals based on the Euclidean distance and the adaptive threshold of the optimal weight.

[0225] In one implementation, when using the radar signal open set identification method described in the above steps to identify unknown radar signals, the test set tf_t obtained in S12 can be processed by the trained network and the UAMP dimensionality reduction algorithm to obtain low-dimensional features {y1,…,y it ,…,y It} Calculate the feature vectors to the reliable class centers The distance between the low-dimensional feature y and the low-dimensional feature y it Corresponding adaptive threshold The comparison process enables the identification of unknown radar signals. The specific process is as follows:

[0226] S171: Input tf_t into the network model obtained in S13 to obtain the embedding vector distribution E. test :

[0227] E test ={e1,e2,…,e nt ,…,e Nt}

[0228] In the formula, nt = 1, 2, ..., Nt, e nt Let t be the nt-th embedding vector.

[0229] S172: Embedded vector distribution E test The low-dimensional vector distribution Y is obtained by dimensionality reduction using UMAP. test :

[0230] Y test ={y1,…,y it ,…,y It}

[0231] In the formula, it = 1, 2, ..., It, y it Let be the it-th low-dimensional vector.

[0232] S173: Calculate vector y it To the Reliability Center European distance

[0233]

[0234] S174: Samples that satisfy the following formula are classified as unknown signals:

[0235]

[0236] In the formula, Let y be the it-th sample vector. it With any known class vector center The minimum value of the Euclidean distance.

[0237] S175: The identification of unknown radar signals is ultimately achieved by using metric learning based on adaptive thresholds.

[0238] As can be seen from the above, the embodiments of this application provide a radar signal open set recognition method, which solves the problem of low recognition rate and poor generalization caused by poor selection of empirical thresholds when recognizing radar with low probability of intercept. A reliable sample library is constructed by selecting reliable samples through a sample screening algorithm. An adaptive threshold is then built based on the radar signal sample library to improve the recognition rate and enhance the universality of the radar signal open set recognition method. Simultaneously, an adaptive seabird algorithm is designed to optimize the adaptive threshold weights, constructing an exploration direction factor. Combining spiral optimization search and linear optimization search expands the search range and strengthens the search capability. Adaptive weights and adaptive step sizes that change with the number of iterations are added to balance global and local search capabilities, improving the reliability of the radar signal open set recognition method. Furthermore, an LSTM network is designed and trained to extract features and obtain embedding vectors. By comparing the Euclidean distance between vectors with the threshold, the radar modulation signal is recognized, enhancing the network's generalization and adaptability, improving the accuracy and universality of the radar signal open set recognition method, and effectively increasing the accuracy of unknown signal recognition.

[0239] Example 2

[0240] This application demonstrates the effectiveness of the radar signal open set recognition method in Embodiment 1 through simulation experiments.

[0241] Table 1 shows the parameters for generating radar signals of known types, with a sampling frequency of 256MHz and 1024 sampling points. Table 2 compares the fitness values ​​obtained using the Blue Bowerbird algorithm and the adaptive threshold trained by the adaptive seabird algorithm described in Embodiment 1 of this application. Table 3 compares the network recognition rates of the test set under empirical and adaptive thresholds, with 500 repeated trials.

[0242] Table 1

[0243]

[0244]

[0245] Table 2

[0246] Intelligent Algorithm fitness value Satin Blue Gardener Bird Algorithm 0.835 Adaptive Seabird Algorithm 0.996

[0247] Table 3

[0248] Threshold type Recognition rate Adaptive threshold 96％ Experience threshold 91％

[0249] In Embodiment 2 of this application, 1000 known radar signals are generated as a training set through signal modeling, and 1000 data points containing unknown signals are obtained as a test set. The experiment is repeated 500 times, and the method described in Embodiment 1 of this application is used to solve the problem. The neural network is trained using the training set, and the recognition rate is obtained using the test set. Compared to the Bluebird Algorithm, the method in Embodiment 1 of this application has a higher fitness value of 0.996, and the designed network uses an adaptive threshold, achieving an unknown signal recognition rate as high as 96%.

[0250] Therefore, the radar signal open set recognition method based on the adaptive seabird algorithm provided in this application embodiment can not only improve the network convergence speed and increase the network's generalization and reliability, but also accurately recognize unknown radar signals, and has good applicability.

[0251] Example 3

[0252] This application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. The memory stores software programs and modules, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory and processor are connected via a bus. Specifically, the processor implements any of the steps in Embodiment 1 by running the computer program stored in the memory.

[0253] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0254] Memory may include read-only memory, flash memory, and random access memory, and provides instructions and data to the processor. Some or all of the memory may also include non-volatile random access memory.

[0255] As can be seen from the above, the electronic device provided in this application embodiment, by running a computer program to implement the radar signal open set identification method as described in Embodiment 1 above, can not only improve the network convergence speed and increase the network's generalization and reliability, but also accurately identify unknown radar signals, thus possessing good applicability.

[0256] It should be understood that if the integrated modules / units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0257] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0258] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0259] It should be noted that the methods and detailed examples provided in the above embodiments can be incorporated into the apparatus and devices provided in the embodiments, and can be referred to each other, without further elaboration.

[0260] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0261] In the embodiments provided in this application, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of the modules or units described above is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0262] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A radar signal open set identification method based on adaptive seabird algorithm, characterized in that, The method comprises the following steps: obtaining a training set, a validation set and a test set, wherein the training set comprises a plurality of known radar signals, and the validation set and the test set each comprise a plurality of unknown radar signals; constructing an LSTM network, training the LSTM network based on the training set, and obtaining a feature extraction network; obtaining embedding vectors of the training set and the validation set respectively according to the feature extraction network, and performing dimension reduction processing on the embedding vectors of the training set and the validation set respectively by using a UMAP algorithm to obtain corresponding low-dimensional features; screening reliable samples from the low-dimensional features of the training set by using a distance-based sample screening algorithm to construct a known radar signal sample library; constructing an adaptive threshold for unknown radar signal identification according to the low-dimensional features of the sample library and the validation set; designing an adaptive seabird algorithm to optimize the weight of the adaptive threshold, combining spiral optimization search and straight line optimization search, designing adaptive weight and adaptive step size, and introducing an exploration direction factor for iterative calculation to obtain an optimal weight; identifying unknown radar signals in the test set according to the feature extraction network and the adaptive threshold based on the optimal weight. The method of designing an adaptive seabird algorithm to optimize the weight of the adaptive threshold, combining spiral optimization search and straight line optimization search, designing adaptive weight and adaptive step size, and introducing an exploration direction factor for iterative calculation to obtain an optimal weight comprises the following steps: randomly generating an initial population comprising a plurality of courting pavilions, and iteratively updating the initial population by repeatedly performing the following steps until a global optimal solution is obtained as the optimal weight: calculating the initial positions of the courting pavilions, and the attraction and fitness values of the courting pavilions; combining spiral optimization search and straight line optimization search to calculate the potential positions of the courting pavilions; designing adaptive step size and adaptive weight respectively, and adding an exploration direction factor for maintaining population diversity to obtain new fitness values corresponding to the potential positions; comparing the fitness values with the new fitness values, updating the positions of the courting pavilions, and randomly selecting a dimension for position mutation update for part of the courting pavilions in each courting pavilion; combining the courting pavilions obtained by position update and the courting pavilions obtained by position mutation update to obtain a combined population, sorting the cost function values of all courting pavilions in the combined population from small to large, retaining courting pavilions whose cost function values meet a preset condition, and eliminating other courting pavilions.

2. The radar signal open set identification method of claim 1, wherein, The method of training the LSTM network based on the training set and obtaining a feature extraction network comprises the following steps: performing wavelet scattering transformation on the known radar signals in the training set and the unknown radar signals in the validation set respectively to obtain scattering features of the training set and the validation set; training the LSTM network by using the scattering features of the training set as training data to obtain the feature extraction network, wherein the feature extraction network comprises an input layer, an LSTM layer, a first full connection layer, a second full connection layer and a softmax layer.

3. The radar signal open set identification method of claim 2, wherein, The obtaining of the embedding vectors of the training set and the validation set respectively according to the feature extraction network comprises: The following steps are performed on the scattering features of the training set and the scattering features of the validation set respectively: The corresponding scattering features are input into the input layer for initialization operation to obtain the 0th layer output, the 0th layer output is input into the LSTM layer for calculation to obtain the 1st layer output, the 1st layer output is input into the first full connection layer for calculation to obtain the 2nd layer output, the 2nd layer output is input into the second full connection layer for calculation to obtain the 3rd layer output, and the 3rd layer output is taken as the corresponding embedding vector.

4. The radar signal open set identification method of claim 1, wherein, The dimension reduction processing of the embedding vectors of the training set and the validation set respectively by the UMAP algorithm comprises: The following steps are performed on the embedding vectors of the training set and the embedding vectors of the validation set respectively: According to the embedding vectors, the nearest neighbor algorithm is used to obtain the adjacent data; According to the adjacent data, the joint probability and the low-dimensional space distribution of the joint probability are calculated, the cross entropy of the joint probability and the low-dimensional space distribution of the joint probability is calculated as a loss function; According to the loss function, a dimension reduction model is obtained by a gradient descent algorithm; According to the dimension reduction model, the embedding vectors are mapped to a low-dimensional feature space to obtain corresponding low-dimensional features.

5. The radar signal open set identification method of claim 1, wherein, The reliable samples are screened out by the distance-based sample screening algorithm to construct a known radar signal sample library, which comprises: The low-dimensional features of the training set are divided into k sets according to the labels, and k is the number of signal types; The inter-class distance of each set is calculated to obtain a distance matrix; Based on the distance matrix, the distance-based sample screening algorithm is used to screen out reliable samples of each signal type as the known radar signal sample library.

6. The radar signal open set identification method of claim 1, wherein, The construction of the adaptive threshold for unknown radar signal recognition comprises: The class center of the reliable samples is calculated; The Euclidean distance from the low-dimensional features of the validation set to the class center is calculated; The adaptive threshold is constructed according to the Euclidean distance.

7. The radar signal open set identification method of claim 6, wherein, The unknown radar signal recognition of the test set according to the feature extraction network and the adaptive threshold based on the optimal weight comprises: The embedding vectors of the test set are obtained according to the feature extraction network; The embedding vectors of the test set are dimensionally reduced by the UMAP algorithm to obtain low-dimensional features of the test set; The Euclidean distance from the low-dimensional features of the test set to the class center is calculated; The unknown radar signal recognition of the test set is performed according to the Euclidean distance and the adaptive threshold based on the optimal weight.

8. An electronic device comprising: The memory, the processor and the computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the steps of the method according to any one of claims 1 to 7.

Images