A new unknown radio source signal recognition method based on a rolling tail honey bear algorithm

Abstract

The application discloses a new unknown radiation source signal identification method based on a ring-tailed honey bear algorithm, comprising: performing Cui-Williams distribution processing on information in a data set of an original radiation source modulation signal to obtain a time-frequency graph; training a preset initial identification model through the time-frequency graph to obtain an initial discrimination result; designing a ring-tailed honey bear algorithm according to the initial discrimination result, optimizing the trained model to obtain an optimal identification model; and using the optimal identification model to obtain an unknown radiation source signal identification result. The application builds a derivative convolutional autoencoder, designs a derivative loss, constructs a discrimination model fusion mode, fuses discrimination results of each model, increases the robustness of unknown radiation source discrimination, designs a ring-tailed honey bear algorithm, constructs an aroma factor, adds a jumping attribute to the ring-tailed honey bear, optimizes the position of the ring-tailed honey bear, makes the discrimination model threshold constantly improved with the number of iterations, improves the discrimination ability of the model, and enhances the generalization and accuracy of the unknown radiation source signal identification method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of signal recognition, and particularly relates to a new unknown radiation source signal recognition method based on a swallow-tailed bear algorithm. BACKGROUND

[0002] Radiation source modulation recognition is an important link in the field of electronic reconnaissance. With the increasingly complex electromagnetic environment, new radar systems are constantly emerging, and the received radiation source signals have modulation types that do not exist in the radar library, which leads to the inability of the modulation recognition model to recognize, bringing new challenges to radiation source modulation recognition. Distinguishing unknown radiation source signals from received signals is a prerequisite for improving the recognition rate of radiation source modulation signals. Therefore, how to accurately recognize unknown radiation source signals plays an important role in improving reconnaissance capabilities.

[0003] At present, the recognition methods for unknown radiation source signals include a clustering-based recognition method and an anomaly detection-based recognition method. The clustering-based recognition method has a fuzzy clustering boundary, which leads to limited recognition performance. The anomaly detection-based recognition method mainly includes statistical, distance, density-based methods, etc. The above methods need to rely on expert experience to set parameters, and have the problem of insufficient generalization.

[0004] Through the search of the prior art documents, it is found that Yang Zhiyuan builds a CNN model for feature extraction in the paper "Radar radiation source signal open set recognition method based on CNN", and uses confidence score to measure the difference between known class signals and unknown class signals, which improves the recognition accuracy of unknown radiation source signals. However, due to the great similarity of radiation source signal time-frequency graphs, this method is easy to divide unknown class samples into known classes, and the recognition ability has room for improvement. Liu Huiling builds a deep autoencoder in the paper "Radar radiation source recognition and behavior cognition based on deep learning", measures the reconstruction error size and feature similarity between the original sample and the reconstructed sample to distinguish known and unknown radiation source signals, and improves the recognition performance by setting two measurement standards. However, since the reconstruction error threshold and the feature similarity threshold depend on human setting, the generalization is poor, and it is difficult to achieve reliable detection. There is also a "Radar signal unknown modulation method detection method based on generative adversarial network", which introduces the crow search algorithm into the generative adversarial network model to realize adaptive setting of the detection threshold and improve the generalization of the detection algorithm. However, only a single-class recognition network is used, which has the problems of poor robustness and poor recognition accuracy. If multiple-class detection models are combined and further optimized by swarm intelligence algorithms, the robustness and generalization of the detection model can be effectively improved, and accurate recognition of unknown radiation sources can be realized. SUMMARY

[0005] To solve the above technical problems, the application provides a new unknown radiation source signal identification method based on a honey badger algorithm, which can effectively improve the robustness and generalization of the detection model, and realize accurate identification of unknown radiation sources.

[0006] To achieve the above purpose, the application provides a new unknown radiation source signal identification method based on a honey badger algorithm, which includes:

[0007] A data set of original radiation source modulation signals is constructed, and the information in the data set is processed by Cui-Williams distribution to obtain a time-frequency graph;

[0008] The preset initial identification model is trained through the time-frequency graph, and an initial discrimination result is obtained;

[0009] According to the initial discrimination result, a honey badger algorithm is designed to optimize the trained preset initial identification model and obtain an optimal identification model;

[0010] The optimal identification model is used to obtain an unknown radiation source signal identification result.

[0011] Optionally, the preset initial identification model includes a ResNet18 sub-model, a derivative convolutional autoencoder, and a GANomaly sub-model;

[0012] Obtaining the initial discrimination result includes:

[0013] The time-frequency graph is input into the ResNet18 sub-model to obtain a first discrimination result;

[0014] The time-frequency graph is input into the derivative convolutional autoencoder to obtain a second discrimination result;

[0015] The time-frequency graph is input into the GANomaly sub-model to obtain a third discrimination result;

[0016] The first discrimination result, the second discrimination result, and the third discrimination result are fused, if the discrimination results of two or more models are all known categories, the output result is a known category, otherwise, the output result is an unknown category.

[0017] Optionally, obtaining the first discrimination result includes:

[0018] The time-frequency graph is input into the ResNet18 sub-model to obtain the distance from each sample to the corresponding category feature center, the distance is used to fit a Weibull distribution, and the scale factor and the position factor of the Weibull distribution are obtained;

[0019] inputting the time-frequency graph into the ResNet18 sub-model again to obtain an output feature, and obtaining a probability value of the time-frequency graph belonging to a known category based on the scale coefficient and the position coefficient and the output feature;

[0020] presetting an initial threshold value of the ResNet18 sub-model, performing known-unknown category division on the time-frequency graph by comparison with the probability value to obtain the first discrimination result.

[0021] Optionally, the derived convolutional autoencoder comprises an encoder and a decoder.

[0022] The encoder comprises four convolutional blocks, each of which comprises one convolutional layer and one pooling layer, and the activation function adopts ReLU.

[0023] The decoder comprises four deconvolutional blocks, each of which comprises one transpose convolutional layer and one up-sampling layer, and the activation function adopts ReLU.

[0024] Optionally, obtaining the second discrimination result comprises:

[0025] S1. inputting the time-frequency graph as the input of the encoder, and obtaining the output result of the encoder through four convolutional blocks;

[0026] S2. inputting the output result of the encoder as the input of the decoder, and obtaining the output result of the decoder through four deconvolutional blocks;

[0027] S3. obtaining a derived loss based on the output result of the decoder and the time-frequency graph;

[0028] S4. training the derived convolutional autoencoder based on the derived loss, updating the weight coefficient and the bias coefficient of the derived convolutional autoencoder by using a back propagation gradient algorithm until the model converges, and obtaining the trained derived convolutional autoencoder;

[0029] S5. inputting the time-frequency graph into the trained derived convolutional autoencoder, repeating steps S1 to S2 to obtain a reconstructed sample, and obtaining the difference between the reconstructed sample and the time-frequency graph;

[0030] S6. presetting an initial threshold value of the derived convolutional autoencoder, performing known-unknown category division on the time-frequency graph by comparison with the difference to obtain the second discrimination result.

[0031] Optionally, obtaining the third discrimination result comprises:

[0032] training the GANomaly sub-model based on the time-frequency graph;

[0033] inputting the time-frequency diagram into the trained GANomaly sub-model again to obtain a hidden vector reconstruction error of the GANomaly sub-model;

[0034] presetting an initial threshold value of the GANomaly sub-model, performing unknown class division on the time-frequency diagram by comparison with the hidden vector reconstruction error to obtain the third discrimination result.

[0035] Optionally, obtaining the best recognition model comprises:

[0036] Step 1. initializing population parameters, position information and exploration factor Ψ of the wobbegong bear; wherein the position of each wobbegong bear comprises: an initial threshold value of the ResNet18 sub-model, an initial threshold value of the derived convolutional autoencoder and an initial threshold value of the GANomaly sub-model;

[0037] Step 2. obtaining the fitness value of the kth wobbegong bear;

[0038] Step 3. updating the global optimal position of the wobbegong bear;

[0039] Step 4. calculating the pheromone factor of each wobbegong bear;

[0040] Step 5. comparing the size of the pheromone factor and the exploration factor, and updating the exploration position of each wobbegong bear based on the comparison result; if the pheromone factor is smaller than the exploration factor, step 6 is executed, otherwise step 7 is executed;

[0041] Step 6. updating the exploration position of each wobbegong bear, and repeating steps 4 to 5;

[0042] Step 7. updating the leader position of the wobbegong bear; sorting the fitness values of each wobbegong bear from small to large, and taking the first half of the wobbegong bear as the leader and the second half of the wobbegong bear as the follower;

[0043] Step 8. updating the position of the follower of the wobbegong bear;

[0044] Step 9. when the prey approaches the wobbegong bear, updating the position of all wobbegong bears again;

[0045] Step 10. recalculating the fitness value of the current wobbegong bear population and updating the global optimal position of the wobbegong bear;

[0046] Step 11. repeating steps 4 to 10 until the iteration number reaches MIter, and obtaining the global optimal solution will and As the final threshold of the ResNet18 submodel, the derived convolutional autoencoder and the GANomaly submodel, based on the final threshold, the best recognition model is obtained.

[0047] Optionally, a new type of class radiation source signal is obtained.

[0048] Mixing the new type of class radiation source signal with the original radiation source modulation signal to construct a new radiation source modulation signal,

[0049] Performing a Cauchy-Williams distribution processing on the new radiation source modulation signal to obtain a new time-frequency graph.

[0050] Inputting the new time-frequency graph into the best recognition model to obtain a final discrimination result, fusing the final discrimination result to obtain an unknown radiation source signal recognition result.

[0051] Compared with the prior art, the present application has the following advantages and technical effects:

[0052] The present application aims to solve the problems of existing unknown radiation source signal recognition algorithms, such as weak robustness, insufficient generalization ability and low recognition ability, and designs a new unknown radiation source signal recognition method based on the convolutional tail honey algorithm. The present application builds a derived convolutional autoencoder, designs a derived loss, constructs a discrimination model fusion method, fuses the discrimination results of each model, and increases the robustness of unknown radiation source discrimination. At the same time, the convolutional tail honey algorithm is designed, the aroma factor is constructed, the jump attribute is added to the convolutional tail honey, the position of the convolutional tail honey is optimized, the discrimination model threshold is constantly improved with the number of iterations, the discrimination ability of the model is improved, and the generalization and accuracy of the unknown radiation source signal recognition method are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are used to explain the illustrative embodiments of the present application and their descriptions, and do not constitute improper limitations on the present application. In the drawings:

[0054] Figure 1 The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are used to explain the illustrative embodiments of the present application and their descriptions, and do not constitute improper limitations on the present application. In the drawings:

[0055] Figure 2 The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are used to explain the illustrative embodiments of the present application and their descriptions, and do not constitute improper limitations on the present application. In the drawings: DETAILED DESCRIPTION

[0056] It should be noted that the embodiments and features in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0057] It is noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0058] The purpose of the embodiment is to design an unknown radiation source signal recognition method, mainly to solve the problems of existing unknown radiation source recognition algorithm, such as weak robust performance, insufficient generalization ability, low recognition ability, etc. The present application proposes a derivative loss, builds a derivative convolutional autoencoder, fuses multiple discriminant models, designs a wapiti algorithm to intelligently set multiple discriminant model thresholds, adds a jump attribute to the wapiti, introduces a fragrance factor, and improves the wapiti update method. The optimized parameters can accurately distinguish unknown radiation source signals, and realize unknown radiation source modulation signal recognition.

[0059] As shown in the accompanying Figure 1 , it is a new wapiti-based unknown radiation source signal recognition method flowchart of the embodiment.

[0060] Step 1: Generate a raw radiation source modulation signal data set x(t) containing six known categories, and make a Choi-Williams distribution on x(t) to get a time-frequency diagram Fig.

[0061] Step 1.1: Generate a raw radiation source modulation signal data set x(t) containing six known categories:

[0062] x(t) = {x_old1(t), x_old2(t), …, x_old c (t), …, x_old6(t)} (1)

[0063] In the formula, t is time, x_old c (t) is the c-th signal set in x(t), and the expression of x_old c (t) is:

[0064]

[0065] In the formula, Num c is the number of c-th signals, is the q-th signal in x_old c (t); is the Num c th signal in x_old c (t);

[0066] Step 1.2: Make a Choi-Williams distribution on x(t) to get a time-frequency diagram Fig of the raw radiation source modulation signal, and Fig is specifically expressed as:

[0067]

[0068] Num all is the total number of samples of x(t), fig m is the mth time-frequency figure of Fig, m = 1, 2, …, Num is the Num all th time-frequency figure of Fig. fig m The calculation formula is as follows:

[0069]

[0070] In the formula, ω is the angular frequency, ε is the scaling factor, τ is the time shift variable, μ is the integral variable, (·) * represents the conjugate operation, e (·) is the exponential operation, e -jωτ is the exponential operation on -jωτ, j represents the imaginary number, dμ is the integral on μ, and dτ is the integral on τ.

[0071] Step 2: Migrate the ResNet18 model, design a derivative convolutional autoencoder, and build a GANomaly model. Use the time-frequency figures Fig generated in step 1 to train the three models respectively. Set the threshold parameters δ tran , δ CAE and δ GAN for the three models respectively, design the fusion method of each model, and obtain the initial discrimination result Result_final.

[0072] Step 2.1: Migrate the pre-trained ResNet18 model model1 and output the discrimination result Result tran of model1 on the time-frequency figure Fig.

[0073] Step 2.1.1: Migrate the ResNet18 network model model1 pre-trained based on the ImageNet image dataset.

[0074] Step 2.1.2: Put the time-frequency figure Fig generated in step 1.2 into model1 in step 2.1.1, and calculate the distance dist from each sample to the corresponding class feature center:

[0075] dist = ||TF, TF_center||2 (5)

[0076] In the formula, ||·||2 represents the L2 norm operation, TF represents the output feature of the time-frequency figure Fig after the 17th convolutional layer in model1, and TF_center is the corresponding class feature center. TF_center is specifically designed as:

[0077] TF_center = {TF_c1, TF_c2, …, TF_c6} (6) c ,…,TF_c6} (6)

[0078] where TF_c c is the feature center of the c-th class. TF_c c is specifically designed as:

[0079]

[0080] where TF_l is the output feature of the l-th sample in the c-th class by model1.

[0081] Step 2.1.3: Send the time-frequency diagram Fig back to model1 in step 2.1.1 to obtain the output feature TF_re, and calculate the probability value P that the time-frequency diagram Fig belongs to the known class:

[0082]

[0083] where p m is the output probability value of the m-th sample fig m , is the output probability value of the Num all -th sample. p m is specifically designed as:

[0084]

[0085] where max(·) is the maximum value operation, TF_re m is the output feature of the m-th sample, ||TF_re m -TF_c c ||2 represents the distance from TF_re m to TF_c c , θ c and λ c are the scale coefficient and position coefficient obtained after Weibull fitting of dist, and F(·) is the Weibull probability distribution function. F(·) is specifically designed as:

[0086]

[0087] Step 2.1.4: Set the threshold value δ tran of model1, and divide the time-frequency diagram Fig into known and unknown classes to obtain the discrimination result Result tran :

[0088]

[0089] where Fig m Discrimination results of the time-frequency map Fig by model 1 Num all Discrimination results of the time-frequency map Fig by model 1 The specific design is as follows:

[0090]

[0091] In the formula, 1 represents a known class, and 0 represents an unknown class.

[0092] Step 2.2: build and train the derived convolutional autoencoder model 2, and output the discrimination result Result of the time-frequency map Fig by model 2 CAE In the designed derived convolutional autoencoder, the encoder is composed of four convolutional blocks, each of which contains one convolutional layer and one pooling layer, and the activation function adopts ReLU. The decoder is composed of four deconvolutional blocks, each of which is composed of one transpose convolutional layer and one up-sampling layer, and the activation function adopts ReLU.

[0093] Step 2.2.1: take the time-frequency map Fig generated in step 1.2 as the input F_encoder of the encoder in model 2 out(0) Repeat the following formula 4 times to obtain the output result F_encoder of the fourth convolutional block out(4) .

[0094]

[0095] In the formula, n e is the number of convolutional blocks, n e = 1, 2, 3, 4, represents the output feature of the time-frequency map Fig by the nth e convolutional block, avgpool(·) is the average pooling operation, R(·) is the ReLU activation function, ∑(·) is the summation operation, and ⊙ represents the convolution operation, represents the weight coefficient of the nth e convolutional layer, represents the bias coefficient of the nth e convolutional layer.

[0096] Step 2.2.2: take the output F_encoder of step 2.2.1 as the input of the decoder in model 2, and repeat the following formula 4 times to obtain the output result F_decoder of the fourth deconvolutional block out(4) , and reconstruct the input image. out(4)

[0097] ​

[0098] where n d is the number of deconvolution blocks, n d = 1, 2, 3, 4, denotes the output feature of the time-frequency map Fig after the n d th deconvolution block, upsample(·) is the up-sampling operation, denotes the transpose convolution operation, denotes the weight coefficient of the n d th deconvolution layer, denotes the bias coefficient of the n d th deconvolution layer.

[0099] Step 2.2.3: Calculate the derived loss L out(4) from the F_decoder CAE obtained in step 2.2.2 and the time-frequency map Fig generated in step 1.2:

[0100] L CAE = (||F_decoder out(4) - Fig||2) 2 (15)

[0101] where (·) 2 is the square operation.

[0102] Step 2.2.4: Train model2, update the weight coefficients and bias coefficients of the derived convolutional autoencoder using the backpropagation gradient algorithm until the model converges, and save the converged model.

[0103] Step 2.2.5: Re-input the time-frequency map Fig obtained in step 1.2 into the trained model2, repeat steps 2.2.1 to 2.2.2 to obtain the reconstructed sample Frec, and calculate the difference Err CAE between the reconstructed sample Frec and the time-frequency map Fig:

[0104]

[0105] where is the reconstruction error of fig m after model2, is the reconstruction error of the Num all th time-frequency map after model2. The specific design is:

[0106]

[0107] where frec m is the mth reconstructed sample in Frec, and ||·||1 is the L1 norm operation.

[0108] Step 2.2.6: Set the threshold value δ of model2 CAE , and the classification result Result is obtained CAE :

[0109]

[0110] wherein, is fig m the classification result of model2, is the classification result of the Num all th time-frequency graph by model2. The specific design is:

[0111]

[0112] Step 2.3: Construct and train the GANomaly model model3, and output the classification result Result of model3 on the time-frequency graph Fig GAN .

[0113] Step 2.3.1: Put the time-frequency graph Fig generated in step 1.2 into the GANomaly model, train the GANomaly network model, and save the trained GANomaly network model.

[0114] Step 2.3.2: Put the time-frequency graph Fig into the trained GANomaly model again, and calculate the GANomaly latent vector reconstruction error Err GAN :

[0115]

[0116] wherein, is fig m the latent vector reconstruction error of model3, is the latent vector reconstruction error of the Num all th time-frequency graph by model3. The specific design is:

[0117]

[0118] wherein, a_rec m is fig m the output of the encoding network En(·) of the GANomaly model, a m is the output of the encoder G_en(·) in the GANomaly model generator. a m The specific design is:

[0119] am = G_en(fig m ) (22)

[0120] a_rec m Specifically designed as:

[0121] a_rec m = En(fig m ) (23)

[0122] In the formula, fgan m is the reconstruction of fig m The time-frequency diagram, a_rec m is the output of fgan m En(·).

[0123] fgan m Specifically designed as:

[0124] fgan m = G_de[G_en(fig m )] (24)

[0125] In the formula, G_de(·) is the decoder model in the generator.

[0126] Step 2.3.3: Set the threshold value δ of model3 GAN , the time-frequency diagram Fig is divided into unknown categories, and the discrimination result Result GAN :

[0127]

[0128] In the formula, is the discrimination result of fig m by model3, is the discrimination result of the Num all th time-frequency diagram by model3. Specifically designed as:

[0129]

[0130] Step 2.4: Design the discrimination model fusion method, fuse the discrimination results of steps 2.1 to 2.3, if two or more model discrimination results are known categories, the output result is known category, otherwise, the output result is unknown category. Get the initial discrimination result Result_final of the time-frequency diagram Fig:

[0131]

[0132] In the formula, Result mThe final discrimination result of the mth sample, The final discrimination result of the Num all th sample. Result m The specific design is:

[0133]

[0134] In the formula, S(·) is a discrimination function, and the specific design is:

[0135]

[0136] In the formula, h represents the input of S(·).

[0137] Step 3: Combine the initial discrimination result to construct an adaptive function, design a wobbly bear algorithm to optimize the threshold values δ forest , δ CAE and δ GAN of each model in step 2. The wobbly bear algorithm is based on the long-nosed bear algorithm, which gives each long-nosed bear a jumping attribute, introduces a fragrance factor, and optimizes the position update mechanism of the individual to obtain the optimal solution. The output optimal solution is used as the threshold value of each model in step 2 to obtain the best migration model Best derived convolutional autoencoder and best GANomaly model

[0138] As shown in the accompanying Figure 2 , it is a wobbly bear algorithm schematic diagram of the embodiment of the present application.

[0139] Step 3.1: Initialize the population parameters, position information and exploration factor Ψ of the wobbly bear. In this embodiment, the position of each wobbly bear is composed of the threshold values δ tran , δ CAE and δ GAN of the three discrimination models, and the exploration factor Ψ∈(0,1]. The initial position of the kth wobbly bear is:

[0140] MC k = MC_l+η1·(MC_u-MC_l) (30)

[0141] In the formula, MC k is the initial value of the kth wobbly bear, k=1,2,…,K, K is the population size, set K / 2 leaders and K / 2 followers, MC_l is the lower bound of the parameter range, MC_u is the upper bound of the parameter range, and η1 is a random number in [0,1].

[0142] Step 3.2: Calculate the fitness value Fit k of the kth wobbly bear:

[0143] Fit k = mean(Result_final k )-1 (31)

[0144] where mean(·) is the mean operation, Result_final k is the discrimination result of the kth raccoon dog.

[0145] Step 3.3: Update the global optimal position of the raccoon dog:

[0146]

[0147] where MC max is the individual with the maximum fitness value in the current population, and MC best is the global optimal individual.

[0148] Step 3.4: Calculate the aroma factor γ of each raccoon dog:

[0149]

[0150] where η2 is a random number in [0, 1], MC k (iter) is the position information of the kth raccoon dog at the iterth iteration, iter is the current iteration number, MC best (iter) is the position information of the global optimal individual at the iterth iteration, and κ is an adjustment factor, κ ∈ (0, 0.1].

[0151] Step 3.5: Compare the size of γ and Ψ, and update the exploration position of each raccoon dog. If γ < Ψ, execute step 3.6; if γ ≥ Ψ, execute step 3.7.

[0152] Step 3.6: Update the exploration position of each raccoon dog, and repeat steps 3.4 to 3.5. The exploration position update formula is specifically designed as:

[0153] MC k (iter+1) = MC k (iter) + G·[(MC_u-MC_l)·η3+MC_l] (34)

[0154] where η3 is a random number in [0, 1], and G is an exploration factor, which is specifically designed as:

[0155]

[0156] In the formula, Λ is a position coefficient, Λ ∈ (0, 5], ξ is an environment coefficient, ξ ∈ (0, 5], ρ is a shape coefficient, ρ ∈ (0, 5], and MIter is the maximum number of iterations. In the first embodiment, Λ is 1.0, ξ is 3.5, and ρ is 3.0.

[0157] Step 3.7: Update the position of the honey bear leader. The fitness values of each honey bear are sorted from small to large, and the first half of the honey bears are taken as leaders, and the last half of the honey bears are taken as followers. The position of the honey bear leader is updated as follows:

[0158]

[0159] In the formula, is the position information of the u-th leader at the (iter+1)-th iteration, u = 1, 2, …, K / 2, g is the acceleration of gravity, g = 9.81, η4 is a random number in [0, 1], is the velocity information of the u-th leader at the (iter+1)-th iteration, and the calculation formula is:

[0160]

[0161] In the formula, η5 and η6 are both random numbers in [0, 1], is the leader historical control factor, N is the parameter dimension, LE u is the leader best control factor.

[0162] The specific design is as follows:

[0163]

[0164] In the formula, is the historical best position of the n-th dimension parameter of the u-th leader, is the n-th dimension parameter information of the u-th leader at the iter-th iteration.

[0165] LE u The specific design is as follows:

[0166]

[0167] Step 3.8: Update the position of the honey bear follower:

[0168]

[0169] In the formula, is the position information of the w-th follower at the (iter+1)-th iteration, w = 1, 2, …, K / 2. η7 is a random number in [0, 1], FE w is the follower best control factor, Fitw fitness value of the wth follower, Fit best fitness value of the global optimal individual, J(·) is the sign function, which controls the direction of the waddies. J(·) is designed as:

[0170]

[0171] where b is the input of J(·).

[0172] FE w is designed as:

[0173]

[0174] Step 3.9: When the prey approaches the waddies, update the positions of all the waddies again:

[0175]

[0176] where η8is a random number in [0, 1], and JBdenotes the escape factor. JBis designed as:

[0177] JB= 1 - 2 - cos(2 - π - η9) (44)

[0178] where η9is a random number in [0, 1].

[0179] Step 3.10: Calculate the fitness value of the current waddie population by the formula, and update the global optimal position of the waddies according to the formula.

[0180] Step 3.11: Repeat steps 3.4 to 3.10 until the number of iterations reaches MIter, and obtain the global optimal solution and as the threshold values of the three models in step 2, and obtain the best migration model the best derived convolutional autoencoder and the best GANomaly model

[0181] Step 4: Generate a new radiation source modulation signal x containing ten modulation modes of new types and known types new (t), in the same way as generating the time-frequency graph in step 1.2, generate the time-frequency graph Fig new (t) of x new . Send Fig new to the and ​In the middle, the corresponding model discriminant result is obtained, and the discriminant result is obtained by the discriminant model fusion mode designed in step 2.4.

[0182] Step 4.1: Generate four new types of radiation source signals, generate six known types of radiation source signals according to step 1.1, mix the new types and known types of radiation source signals to form new radiation source modulation signals x new (t):

[0183] x new (t)={x_new1(t),x_new2(t),…,x_new υ (t),…,x_new 10 (t)} (45)

[0184] In the formula, x_new υ (t) is the υth signal set in x new (t), and the expression of x_new υ (t) is:

[0185]

[0186] In the formula, is the number of the υth signal, is the yth signal in x_new υ (t), is the yth signal in x_new υ (t).

[0187] Step 4.2: Generate the time-frequency graph Fig new (t) of x new (t) generated in step 1.2.

[0188] Step 4.3: Fig new is passed through the three models in step 3, respectively, to obtain the migration model discriminant result, the derived convolutional autoencoder discriminant result and the GANomaly discriminant result of Fig new .

[0189] Step 4.4: The discriminant result obtained in 4.3 is brought into step 2.4 to obtain the final discriminant result Recog_result. If Recog_result>0, it is discriminated as a known type, otherwise, it is discriminated as an unknown type.

[0190] ​The above merely provides the preferred embodiments of the present application, and the protection scope of the present application is not limited thereto, and any changes or substitutions within the technical scope disclosed by the present application 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 novel method for identifying signals from unknown radiation sources based on the Curly-tailed Honey Bear algorithm, characterized in that, include: A dataset of the original radiation source modulation signal is constructed, and the information in the dataset is processed by the Cui-Williams distribution to obtain the time-frequency diagram; The preset initial identification model is trained using the time-frequency graph, and the initial discrimination result is obtained. Based on the initial discrimination results, a curly-tailed honey bear algorithm is designed to optimize the trained preset initial recognition model and obtain the optimal recognition model. The curly-tailed honey bear algorithm is based on the long-nosed raccoon algorithm, assigning a jumping attribute to each long-nosed raccoon, introducing an aroma factor, and optimizing the individual position update mechanism to obtain the optimal solution. The aroma factor is: in, As aroma factors, for random numbers, For the first Only the curly-tailed honey bear in the first Position information at the next iteration This represents the current iteration number. For the globally optimal individual in the th... Position information at the next iteration As a regulating factor; Using the aforementioned optimal identification model, the identification results of unknown radiation source signals are obtained.

2. The novel method for identifying unknown radiation source signals based on the Curly-tailed Honey Bear algorithm according to claim 1, characterized in that, The preset initial recognition model includes: a ResNet18 sub-model, a derived convolutional autoencoder, and a GANomaly sub-model; Obtaining the initial discrimination result includes: Input the time-frequency graph into the ResNet18 sub-model to obtain the first discrimination result; The time-frequency graph is input into the derived convolutional autoencoder to obtain the second discrimination result; The time-frequency graph is input into the GANomaly sub-model to obtain the third discrimination result; The first, second, and third discrimination results are fused together. If two or more model discrimination results are both known categories, the output result is the known category; otherwise, the output result is the unknown category.

3. The novel method for identifying unknown radiation source signals based on the Curly-tailed Honey Bear algorithm according to claim 2, characterized in that, Obtaining the first discrimination result includes: The time-frequency graph is input into the ResNet18 sub-model to obtain the distance from each sample to the corresponding category feature center. The distance is then used to fit a Weibull distribution to obtain the Weibull distribution's scale factor and position factor. The time-frequency graph is re-input into the ResNet18 sub-model to obtain output features. Based on the scaling factor, position factor, and output features, the probability value of the time-frequency graph belonging to a known category is obtained. An initial threshold is preset for the ResNet18 sub-model. By comparing it with the probability value, the time-frequency graph is classified into unknown categories to obtain the first discrimination result.

4. The novel method for identifying unknown radiation source signals based on the curltail honey bear algorithm according to claim 2, characterized in that, The derived convolutional autoencoder includes: an encoder and a decoder; The encoder includes: 4 convolutional blocks, each convolutional block containing 1 convolutional layer and 1 pooling layer, with ReLU as the activation function; The decoder includes four deconvolutional blocks, each of which includes one transposed convolutional layer and one upsampling layer, with ReLU as the activation function.

5. The novel method for identifying unknown radiation source signals based on the curled-tail honey bear algorithm according to claim 4, characterized in that, Obtaining the second discrimination result includes: S1. Using the time-frequency graph as the input to the encoder, and passing it through 4 convolutional blocks, obtain the output result of the encoder; S2. The output of the encoder is used as the input of the decoder, and after passing through 4 deconvolution blocks, the output of the decoder is obtained; S3. Based on the output of the decoder and the time-frequency diagram, obtain the derived loss; S4. Based on the derived loss, train the derived convolutional autoencoder, and update the weight coefficients and bias coefficients of the derived convolutional autoencoder using the backpropagation gradient algorithm until the model converges, and obtain the trained derived convolutional autoencoder. S5. Input the time-frequency graph into the trained derived convolutional autoencoder, repeat steps S1 to S2, obtain reconstructed samples, and obtain the difference between the reconstructed samples and the time-frequency graph; S6. Preset the initial threshold of the derived convolutional autoencoder, and classify the time-frequency graph into unknown categories by comparing it with the difference, and obtain the second discrimination result.

6. The novel method for identifying unknown radiation source signals based on the curltail honey bear algorithm according to claim 2, characterized in that, Obtaining the third discrimination result includes: The GANomaly sub-model is trained based on the time-frequency graph; The time-frequency graph is re-input into the trained GANomaly sub-model to obtain the latent vector reconstruction error of the GANomaly sub-model. An initial threshold is preset for the GANomaly sub-model. By comparing the threshold with the latent vector reconstruction error, the time-frequency graph is classified into unknown categories to obtain the third discrimination result.

7. The novel method for identifying unknown radiation source signals based on the curltail honey bear algorithm according to claim 2, characterized in that, Obtaining the optimal recognition model includes: Step 1. Initialize the population parameters, location information, and exploration factors of the drongo-tailed honey bear. The position of each humming bear includes: the initial threshold of the ResNet18 sub-model, the initial threshold of the derived convolutional autoencoder, and the initial threshold of the GANomaly sub-model. Step 2. Obtain the first Fitness value for only the Curly-tailed Honey Bear; Step 3. Update the global optimal position of the curly-tailed honey bear; Step 4. Calculate the aroma factor of each capuchin honey bear; Step 5. Compare the aroma factor with the exploration factor, and update the exploration position of each succulent bear based on the comparison result; if the aroma factor is less than the exploration factor, proceed to step 6, otherwise proceed to step 7; Step 6. Update the exploration locations of each Curly-tailed Honey Bear, and repeat steps 4 and 5; Step 7. Update the leader position of the Curly-tailed Honey Bears; sort the fitness values ​​of each Curly-tailed Honey Bear from smallest to largest, and make the first half of the Curly-tailed Honey Bears the leaders and the second half the followers. Step 8. Update the position of the Curly-tailed Honey Bear follower; Step 9. When prey approaches the tango-tailed honey bear, update the location of all tango-tailed honey bears again; Step 10. Recalculate the fitness value of the current tufted-tailed honey bear population and update the global optimum position of the tufted-tailed honey bear; Step 11. Repeat steps 4 through 10 until the number of iterations reaches the target. To obtain the global optimal solution , , ,Will , and The final threshold is used as the final threshold for the ResNet18 sub-model, the derived convolutional autoencoder, and the GANomaly sub-model; based on the final threshold, the best recognition model is obtained.

8. The novel method for identifying unknown radiation source signals based on the curltail honey bear algorithm according to claim 1, characterized in that, Acquire signals from novel types of radiation sources; The novel type of radiation source signal is mixed with the original radiation source modulation signal to construct a new radiation source modulation signal; The new radiation source modulation signal is processed using a Cui-Williams distribution to obtain a new time-frequency diagram; The new time-frequency diagram is input into the optimal identification model to obtain the final discrimination result. The final discrimination result is then fused to obtain the identification result of the unknown radiation source signal.

Images