A radiation source working mode recognition method, program, device and storage medium

Abstract

The present invention belongs to the field of radiation source information identification technology, and specifically relates to a radiation source working mode identification method, program, device and storage medium. The present invention uses cosine metric distance to improve the Gaussian kernel function, improve the model's ability to extract local characteristics, and combines the global ability of the polynomial kernel function to construct a new combined kernel function, improve the SVM model, and improve the robustness of the model. The present invention designs a pygmy hippopotamus algorithm, adaptively optimizes the SVM model parameter combination, gives the pygmy hippopotamus a vigilant attribute, introduces an early warning factor, improves the position update method of the pygmy hippopotamus during defense, and enhances the generalization ability of the algorithm, so that the improved SVM model parameter combination is continuously optimized with the number of iterations, thereby improving the recognition ability of the improved SVM model, and realizing the improvement of the generalization and accuracy of the radiation source working mode identification method under small sample conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of radiation source information recognition, and specifically relates to a radiation source working mode recognition method, program, device and storage medium. Background Art

[0002] Radar emitter working mode recognition is one of the important technologies of electronic countermeasures, providing a basis for subsequent threat level assessment, interference decision-making and radar status prediction. With the complexity and dynamic change of modern electromagnetic environment and the difficulty in obtaining the number of emitter working mode samples due to non-cooperative confrontation, electronic reconnaissance has been greatly challenged.

[0003] At present, there are two main ways to identify the working mode of radiation sources under small sample conditions: the radiation source working mode recognition method based on parameter range matching and the radiation source working mode recognition method based on neural network. The radiation source working mode recognition method based on parameter range matching is to identify and cluster according to the characteristic measurement results. If there is overlap in the working parameters of different modes, different modes will be clustered together, and there will be a problem of fuzzy parameter range. The radiation source working mode recognition method based on neural network increases the difficulty of radar working mode recognition under small sample conditions due to the lack of sufficient samples to support neural network training.

[0004] A search of the existing technical literature found that Zhao Guolin et al. invented a radar working mode recognition method in a patent (patent number: CN201610729293.6), which used four single kernel functions to build an SVM classification model, improving the accuracy of traditional radar working mode recognition. However, since the single kernel function limits the ability to express data and the single kernel function is sensitive to the selection of hyperparameters, this method is prone to misclassifying working mode samples, and the robustness has room for improvement. Chen Xinyu used the simulated annealing algorithm to optimize the parameters of the SVM model in the paper "Research on MFR Working Mode Recognition Based on Machine Learning", which improved the robustness of the traditional SVM model. However, since the selected simulated annealing algorithm is sensitive to the selection of parameters and has poor generalization, it is difficult to achieve reliable recognition. If the SVM model improved by the combined kernel function is used, and the parameters of the improved SVM model are optimized in combination with the Dwarf Hippo algorithm, the robustness and generalization of the detection model can be effectively improved, and the accurate recognition of the working mode of the radiation source under small sample conditions can be achieved. Summary of the invention

[0005] The purpose of the present invention is to design a new radiation source working mode recognition method to address the problems of insufficient robustness, poor generalization, low recognition ability and the like in existing methods when recognizing radiation source working modes under small sample conditions.

[0006] A method for identifying a radiation source working mode comprises the following steps:

[0007] Step 1: Obtain a training set consisting of samples containing pulse description words and radiation source working modes, perform data conversion on the samples in the training set, and generate a two-dimensional image;

[0008] Step 2: Pre-train the feature extraction model to extract features from the two-dimensional image and obtain a feature vector;

[0009] Step 3: Build an SVM classification model and design the kernel function of the SVM model by combining the penalty factor, scaling, polynomial order, and weight coefficient;

[0010] Step 4: Use the Pygmy Hippo algorithm to train the SVM classification model using samples containing feature vectors and radiation source working modes in the training set, and optimize the penalty factor, scaling, polynomial order, and weight coefficient of the SVM classification model;

[0011] Step 5: Obtain the pulse description word of the radiation source to be identified, obtain the corresponding two-dimensional image through data conversion, and obtain the corresponding feature vector through the pre-trained feature extraction model; input the feature vector into the trained SVM classification model to obtain the recognition result of the radiation source working mode.

[0012] Furthermore, the step 4 is specifically as follows:

[0013] Step 4.1: Initialize the number of iterations t = 1, set the maximum number of iterations T_Maxiter, and the population size N of the pygmy hippopotamus; each pygmy hippopotamus individual is composed of the penalty factor, telescoping scale, polynomial order, and weight coefficient of the SVM model; initialize the positions of N pygmy hippopotamus individuals, and calculate the fitness Fit corresponding to the position of each pygmy hippopotamus individual z , take the fitness Fit z The position of the largest Pygmy Hippopotamus individual is taken as the initial global optimal position Dhippo(0);

[0014] Initialize the positions of N Pygmy Hippopotamus individuals as:

[0015] Hippo z (1) = [q z1 ,q z2 ,q z3 ,q z4 ]

[0016] q zm = l m +η1(u m -l m )

[0017] Among them, q z1 ,q z2 ,q z3 ,q z4They represent the four parameters of the zth pygmy hippopotamus, l m Indicates the lower bound of the corresponding parameter, u m The upper bound of the corresponding parameter value is m=1,2,3,4; η1 is a random number in [0,1];

[0018] The fitness Fit z To: Set the position of the zth pygmy hippopotamus individual Hippo z The corresponding penalty factor, scaling scale, polynomial order and weight coefficient are substituted into the SVM model, and the recognition accuracy of the SVM model is obtained through the training set Fit z ;

[0019] Step 4.2: Enter the exploration phase, record the first N2 pygmy hippopotamuses among the N pygmy hippopotamuses as male pygmy hippopotamuses, and record the last N2 pygmy hippopotamuses as female pygmy hippopotamuses, and update the position of each pygmy hippopotamus in the exploration phase according to gender;

[0020]

[0021] in, is the position of the z1th male pygmy hippopotamus in the exploration phase of the tth iteration; is the position of the z2th female pygmy hippopotamus in the exploration phase of the tth iteration; G is a random number in [0,1]; Q is a random number that follows a normal distribution, η2 is a random number in [0,1], η3 is a random number in (0,1], and rand(1,2) is a random number in [1,2]; h1 is an environmental factor, h1=η4+Q, η4 is a random number in [0,1]; T is a distance factor,

[0022] Step 4.3: Calculate the fitness of each male and female pygmy hippopotamus according to their positions during the exploration phase. If the maximum fitness among them is greater than the corresponding fitness of the global optimal position Dhippo(t-1), then update the global optimal position.

[0023] Step 4.4: Enter the predation stage and generate predators that can threaten the survival of the pygmy hippopotamus. The position of the predator in the search space can be expressed as:

[0024] Predator(t)=[p t1 ,p t2 ,p t3 ,p t4 ]

[0025] p tm = l m +η5(u m -l m )

[0026] Among them, η5 is a random number in [0,1];

[0027] Step 4.5: Calculate the fitness of the predator's position Predator(t), according to the warning factor vi of each pygmy hippopotamus at the previous moment z (t-1) Update the position of each pygmy hippopotamus in the population;

[0028]

[0029] Where RL is a random variable with Levy distribution, η6 is a uniform random number in [2,4], η7 is a uniform random number in [2,3], η8 is a uniform random number in [-1,1], η9 is a random number in [0,1], D is the distance between the predator's position and the global optimal position, D = || Predator(t) - Dhippo(t-1)||; Fit Predator(t) <Fit Dhippo(t-1) It means that the fitness corresponding to the predator's position Predator(t) is less than the fitness corresponding to the global optimal position Dhippo(t-1);

[0030] Step 4.6: Update the early warning factors of each pygmy hippopotamus;

[0031]

[0032] Step 4.7: Update the position of each hippopotamus according to the Hippo z (t+1) Calculate fitness and update the global optimal position;

[0033]

[0034] Step 4.8: If t<T_Maxiter, set t=t+1 and return to step 4.2 for the next iteration; otherwise, output the global optimal position and use the penalty factor, scaling scale, polynomial order and weight coefficient of the corresponding SVM model to construct the optimal SVM classification model.

[0035] Furthermore, in step 3, the kernel function of the SVM model designed by combining the penalty factor C, the telescoping scale σ, the polynomial order d and the weight coefficient λ is specifically:

[0036] The optimal margin hyperplane y(x) of the SVM model is expressed as:

[0037] y(x)=sgn(ω·δ(x)+b)

[0038] Where ω is the normal vector of the hyperplane; b is the offset of the hyperplane; sgn(·) is the step function; δ(·) is the feature mapping function of the input sample; x is the vector form of the input sample; y is the vector form of the sample category;

[0039] The objective function of SVM model optimization is expressed as:

[0040]

[0041] Among them, ξ j is the relaxation factor, ξ j ≥0, j=1,2,...,n, n is the sample length;

[0042] The constraints for SVM model optimization are expressed as:

[0043] y j (ω T ·δ(x j )+b)≥1-ξ j

[0044] In the formula, x j is the jth element in x; y j is the jth element in y;

[0045] Construct the kernel function of the SVM model:

[0046]

[0047] In the formula, x T is the transpose of x; ||·|| is the modulo operation; exp(·) is the exponential function;

[0048] According to the constructed SVM model kernel function, the normal vector ω and offset b of the hyperplane are expressed as:

[0049]

[0050] in, a j is the first dual variable to be solved, a' e is the second dual variable to be solved, a=[a1,a2,...,a n ] T is the set of solutions that satisfy the constraints.

[0051] Furthermore, in step 1, a training set consisting of samples including pulse description words and radiation source working modes is obtained, specifically:

[0052] Generates pulse description words PDW containing various types of radiation source working modes:

[0053] PDW=[PDW1,PDW2,...,PDW i ,...,PDW num ]

[0054] Among them, PDW iis the pulse description word of the ith pulse, PDW i =[PW i ,BW i ,CF i ,PRI i ], PW i is the pulse width of the ith pulse; BW i is the pulse bandwidth of the ith pulse; CF i is the pulse carrier frequency of the ith pulse; PRI i is the pulse repetition interval of the i-th pulse; i = 1, 2, ..., num, num is the number of pulses;

[0055] Perform data mapping on all parameters in PDW to form the radiation source working mode parameter data set R_data:

[0056]

[0057] Among them, PW_n i is the i-th pulse width PW in PDW i Normalized results, BW_n i is the bandwidth BW of the i-th pulse in PDW i Normalized results, CF_n i is the carrier frequency CF of the i-th pulse in PDW i Normalized results, PRI_n i is the ith pulse repetition interval PRI in PDW i Normalized results, min(·) is the minimum value in the parameter set, and max(·) is the maximum value in the parameter set.

[0058] Furthermore, in step 1, data conversion is performed on the samples in the training set to generate a two-dimensional image, specifically:

[0059] Assume that the image sample consists of four-channel data of A, R, G, and B. The image sample data set Fig_data is represented as:

[0060] Fig_data=[Fig_data1,Fig_data2,...,Fig_data fp ,...,Fig_data num ]

[0061] Among them, Fig_data fp is the fpth image sample, Fig_data fp=cat(data_A,data_R,data_G,data_B), cat(·) is a channel merging function, data_A is the A channel data of the image sample, data_R is the R channel data of the image sample, data_G is the G channel data of the image sample, and data_B is the B channel data of the image sample;

[0062]

[0063] Among them, k is the number of times each column of R_data is replicated;

[0064] Furthermore, the pre-trained feature extraction model in step 2 is specifically: migrating the PlainMamba network model pre-trained based on the ImageNet image dataset, removing the softmax classification layer in the PlainMamba network model, forming a migration network model model1, and taking the two-dimensional image sample as the input of the model model1.

[0065] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-mentioned radiation source working mode recognition method.

[0066] A computer-readable storage medium stores a computer program / instruction, which implements the steps of the above-mentioned radiation source working mode recognition method when executed by a processor.

[0067] A computer program product comprises a computer program / instruction, which implements the steps of the above-mentioned radiation source working mode recognition method when executed by a processor.

[0068] The beneficial effects of the present invention are:

[0069] The present invention aims at the problems that the existing methods have insufficient robustness, poor generalization, low recognition ability and the like when identifying the working mode of radiation sources under small sample conditions, and provides a new method for identifying the working mode of radiation sources. The present invention uses cosine metric distance to improve the Gaussian kernel function, improves the model's ability to extract local characteristics, and combines the global ability of the polynomial kernel function to construct a new combined kernel function, improve the SVM model, and improve the robustness of the model. At the same time, a pygmy hippopotamus algorithm is designed to adaptively optimize the SVM model parameter combination, give the pygmy hippopotamus a vigilant attribute, introduce an early warning factor, improve the position update method of the pygmy hippopotamus during defense, enhance the generalization ability of the algorithm, and make the improved SVM model parameter combination continuously optimized with the number of iterations, thereby improving the recognition ability of the improved SVM model, and realizing the improvement of the generalization and accuracy of the method for identifying the working mode of radiation sources under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0071] Figure 2 It is a flow chart of the pygmy hippopotamus algorithm in the present invention. DETAILED DESCRIPTION

[0072] The present invention is further described below in conjunction with the accompanying drawings.

[0073] This paper designs a new radiation source working mode recognition method, which mainly solves the problems of insufficient robustness, poor generalization, and low recognition ability in the existing radiation source working mode recognition method under small sample conditions. The present invention uses cosine metric distance to improve the Gaussian kernel function, improve the model's ability to extract local characteristics, and combines the global ability of the polynomial kernel function to construct a new combined kernel function to improve the robustness of the SVM model, design the pygmy hippopotamus algorithm, adaptively optimize the SVM model parameter combination, give the pygmy hippopotamus a vigilant attribute, introduce an early warning factor, and improve the position update method of the pygmy hippopotamus during defense, which enhances the generalization of the algorithm. The optimized parameter combination enables the SVM model to accurately identify the radiation source working mode, and realizes the accurate identification of the radiation source working mode under small sample conditions.

[0074] As attached Figure 1 The figure shows a flow chart of a novel radiation source working mode recognition method of the present invention.

[0075] A method for identifying a radiation source working mode comprises the following steps:

[0076] Step 1: Obtain a training set consisting of samples containing pulse description words and radiation source working modes, perform data conversion on the samples in the training set, and generate a two-dimensional image;

[0077] Step 2: Pre-train the feature extraction model to extract features from the two-dimensional image and obtain a feature vector;

[0078] Step 3: Build an SVM classification model and design the kernel function of the SVM model by combining the penalty factor, scaling, polynomial order, and weight coefficient;

[0079] Step 4: Use the Pygmy Hippo algorithm to train the SVM classification model using samples containing feature vectors and radiation source working modes in the training set, and optimize the penalty factor, scaling, polynomial order, and weight coefficient of the SVM classification model;

[0080] Step 5: Obtain the pulse description word of the radiation source to be identified, obtain the corresponding two-dimensional image through data conversion, and obtain the corresponding feature vector through the pre-trained feature extraction model; input the feature vector into the trained SVM classification model to obtain the recognition result of the radiation source working mode.

[0081] As attached Figure 2 The figure shows a schematic diagram of the pygmy hippopotamus algorithm in the present invention.

[0082] Step 4.1: Initialize the number of iterations t = 1, set the maximum number of iterations T_Maxiter, and the population size N of the pygmy hippopotamus; each pygmy hippopotamus individual is composed of the penalty factor, telescoping scale, polynomial order, and weight coefficient of the SVM model; initialize the positions of N pygmy hippopotamus individuals, and calculate the fitness Fit corresponding to the position of each pygmy hippopotamus individual z , take the fitness Fit z The position of the largest Pygmy Hippopotamus individual is taken as the initial global optimal position Dhippo(0);

[0083] Initialize the positions of N Pygmy Hippopotamus individuals as:

[0084] Hippo z (1) = [q z1 ,q z2 ,q z3 ,q z4 ]

[0085] q zm = l m +η1(u m -l m )

[0086] Among them, q z1 ,q z2 ,q z3 ,q z4 They represent the four parameters of the zth pygmy hippopotamus, l m Indicates the lower bound of the corresponding parameter, u m The upper bound of the corresponding parameter value is m=1,2,3,4; η1 is a random number in [0,1];

[0087] The fitness Fit z To: Set the position of the zth pygmy hippopotamus individual Hippo z The corresponding penalty factor, scaling scale, polynomial order and weight coefficient are substituted into the SVM model, and the recognition accuracy of the SVM model is obtained through the training set Fit z ;

[0088] Step 4.2: Enter the exploration phase, record the first N2 pygmy hippopotamuses among the N pygmy hippopotamuses as male pygmy hippopotamuses, and record the last N2 pygmy hippopotamuses as female pygmy hippopotamuses, and update the position of each pygmy hippopotamus in the exploration phase according to gender;

[0089]

[0090] in, is the position of the z1th male pygmy hippopotamus in the exploration phase of the tth iteration; is the position of the z2th female pygmy hippopotamus in the exploration phase of the tth iteration; G is a random number in [0,1]; Q is a random number that follows a normal distribution, η2 is a random number in [0,1], η3 is a random number in (0,1], and rand(1,2) is a random number in [1,2]; h1 is an environmental factor, h1=η4+Q, η4 is a random number in [0,1]; T is a distance factor,

[0091] Step 4.3: Calculate the fitness of each male and female pygmy hippopotamus according to their positions during the exploration phase. If the maximum fitness among them is greater than the corresponding fitness of the global optimal position Dhippo(t-1), then update the global optimal position.

[0092] Step 4.4: Enter the predation stage and generate predators that can threaten the survival of the pygmy hippopotamus. The position of the predator in the search space can be expressed as:

[0093] Predator(t)=[p t1 ,p t2 ,p t3 ,p t4 ]

[0094] p tm = l m +η5(u m -l m )

[0095] Among them, η5 is a random number in [0,1];

[0096] Step 4.5: Calculate the fitness of the predator's position Predator(t), according to the warning factor vi of each pygmy hippopotamus at the previous moment z (t-1) Update the position of each pygmy hippopotamus in the population;

[0097]

[0098] Where RL is a random variable with Levy distribution, η6 is a uniform random number in [2,4], η7 is a uniform random number in [2,3], η8 is a uniform random number in [-1,1], η9 is a random number in [0,1], D is the distance between the predator's position and the global optimal position, D = || Predator(t) - Dhippo(t-1)||; Fit Predator(t) <Fit Dhippo(t-1) It means that the fitness corresponding to the predator's position Predator(t) is less than the fitness corresponding to the global optimal position Dhippo(t-1);

[0099] Step 4.6: Update the early warning factors of each pygmy hippopotamus;

[0100]

[0101] Step 4.7: Update the position of each hippopotamus according to the Hippo z (t+1) Calculate fitness and update the global optimal position;

[0102]

[0103] Step 4.8: If t<T_Maxiter, set t=t+1 and return to step 4.2 for the next iteration; otherwise, output the global optimal position and use the penalty factor, scaling scale, polynomial order and weight coefficient of the corresponding SVM model to construct the optimal SVM classification model.

[0104] Embodiment 1:

[0105] Step 1: Generate pulse description words PDW containing five categories of radiation source working modes, and perform data mapping processing to form a working mode parameter data set.

[0106] Step 1.1: Generate a pulse description word PDW containing five categories of radiation source working modes:

[0107] PDW=[PDW1,PDW2,...,PDW i ,...,PDW num ]

[0108] Where PDW i is the pulse description word of the i-th pulse, i = 1, 2, ..., num, num is the number of pulses, where PDW i It can be expressed as:

[0109] PDW i =[PW i ,BW i ,CF i ,PRI i ]

[0110] Where PW i is the pulse width of the ith pulse, BW i is the pulse bandwidth of the ith pulse, CF i is the pulse carrier frequency of the ith pulse, PRI i is the pulse repetition interval of the ith pulse.

[0111] Step 1.2: Perform data mapping on all parameters in PDW to form the radiation source working mode parameter data set R_data, which can be expressed as:

[0112]

[0113] Where, PW_n i is the i-th pulse width PW in PDW i Normalized results:

[0114]

[0115] Where min(·) is the minimum value in the parameter set, and max(·) is the maximum value in the parameter set.

[0116] BW_n i is the bandwidth BW of the i-th pulse in PDW i Normalized results:

[0117]

[0118] CF_n i is the carrier frequency CF of the i-th pulse in PDW i Normalized results:

[0119]

[0120] PRI_n i is the ith pulse repetition interval PRI in PDW i Normalized results:

[0121]

[0122] Step 1.3: Perform data conversion on the parameter data set R_data generated in step 1.2 to form the working mode image sample data set Fig_data.

[0123] The image sample is assumed to consist of four-channel data of A, R, G, and B, and contains five working modes. The image sample data set Fig_data can be expressed as:

[0124] Fig_data=[Fig_data1,Fig_data2,...,Fig_data fp ,...,Fig_data num ]

[0125] In the formula, Fig_data fp is the fpth image sample, where Fig_data fp It can be expressed as:

[0126] Fig_data fp =cat(data_A,data_R,data_G,data_B)

[0127] Where cat(·) is the channel merging function, data_A is the A channel data of the image sample, which is specifically expressed as:

[0128]

[0129] Wherein, k is the number of times each column of data in R_data is replicated. In the first embodiment, k is 1024.

[0130] data_R is the R channel data of the image sample, which is specifically expressed as:

[0131]

[0132] data_G is the G channel data of the image sample, which is specifically expressed as:

[0133]

[0134] data_B is the B channel data of the image sample, which is specifically expressed as:

[0135]

[0136] Step 2: Migrate the PlainMamba pre-trained model, use the image sample dataset Fig_data generated in step 1 as the model input, and obtain the working mode feature dataset Feature_data.

[0137] Step 2.1: Migrate the PlainMamba network model pre-trained based on the ImageNet image dataset, remove the softmax classification layer in the PlainMamba network model, and construct the migration network model model1.

[0138] Step 2.2: Take the image sample dataset Fig_data generated in step 1 as the input of model model1, and extract the feature dataset Feature_data, which can be expressed as:

[0139] Feature_data=model1(Fig_data)

[0140] Step 3: Combine the scaling scale σ, polynomial order d and weight coefficient λ to design a new combined kernel function, improve the SVM classification model, use the feature data set Feature_data in step 2 as the input of the improved SVM model, and save the current training model SVM temp (·).

[0141] Step 3.1: The optimal margin hyperplane y(x) of the SVM model can be expressed as:

[0142] y(x)=sgn(ω·δ(x)+b) (32)

[0143] Where ω is the normal vector of the hyperplane, b is the offset of the hyperplane, sgn(·) is the step function, δ(·) is the feature mapping function of the input sample, x is the vector form of input sample 1, and y is the vector form of sample category 1.

[0144] Step 3.2: The objective function of SVM model optimization can be expressed as:

[0145]

[0146] Where C is the penalty factor for misclassified training samples, C∈[4,-4], ξ j is the relaxation factor, ξ j ≥0, j=1,2,...,n, n is the sample length. The constraints for SVM model optimization can be expressed as:

[0147] y j (ω T ·δ(x j )+b)≥1-ξ j (34)

[0148] In the formula, x j is the jth element in x, y j is the jth element in y.

[0149] Step 3.3: Convert equation (33) and equation (34) into the following function:

[0150]

[0151] In the formula, a j The dual variable 1, a' is to be solved e The dual variable 2 to be solved, a=[a1,a2,...,a n ] T To satisfy the solution set under the constraints, x' e is the e-th element of the input sample 2 vector x', y' e is the e-th element of sample category 2y', e = 1, 2, ..., n.

[0152] Step 3.4: Construct a new kernel function for the SVM model:

[0153]

[0154] In the formula, x T is the transpose of x, ||·|| is the modulo operation, and exp(·) is the exponential function.

[0155] Step 3.5: Substitute the kernel function K(x,x') constructed in step 3.4 into the feature map inner product δ(x') of equation (35) in step 3.3 e )δ(x j ), which can be specifically expressed as:

[0156]

[0157] Step 3.6: Solve step 3.5, calculate the hyperplane normal vector ω and the hyperplane offset b in step 3.1, and build an improved SVM model. The normal vector ω of the hyperplane can be expressed as:

[0158]

[0159] The offset b of the hyperplane can be expressed as:

[0160]

[0161] Step 3.7: Use the feature data set Feature_data in step 3.2 as the training input of the improved SVM model and save the current training model SVM temp (·):

[0162] C_result temp =SVM temp (Feature_data) (40)

[0163] Where, C_result temp For the feature data set Feature_data through SVM temp (·) after recognition result.

[0164] Step 4: Use the Pygmy Hippo algorithm to train the SVM classification model using samples containing feature vectors and radiation source working modes in the training set, and optimize the penalty factor, scaling, polynomial order, and weight coefficient of the SVM classification model;

[0165] Step 5: Obtain the pulse description word of the radiation source to be identified, obtain the corresponding two-dimensional image through data conversion, and obtain the corresponding feature vector through the pre-trained feature extraction model; input the feature vector into the trained SVM classification model to obtain the recognition result of the radiation source working mode.

[0166] The present invention uses cosine metric distance to improve Gaussian kernel function, improve the model's ability to extract local characteristics, and combines the global ability of polynomial kernel function to construct a new combined kernel function, improve SVM model, and improve the robustness of the model. At the same time, the pygmy hippopotamus algorithm is designed to adaptively optimize the combination of SVM model parameters, give the pygmy hippopotamus a vigilant attribute, introduce early warning factors, improve the position update method of the pygmy hippopotamus during defense, and enhance the generalization ability of the algorithm, so that the improved SVM model parameter combination is continuously optimized with the number of iterations, thereby improving the recognition ability of the improved SVM model, and realizing the improvement of the generalization and accuracy of the radiation source working mode recognition method under small sample conditions.

[0167] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying a radiation source working mode, characterized in that: The following steps are involved: Step 1: Obtain a training set consisting of samples containing pulse description words and radiation source working modes, perform data conversion on the samples in the training set, and generate a two-dimensional image; Step 2: Pre-train the feature extraction model to extract features from the two-dimensional image and obtain a feature vector; Step 3: Build an SVM classification model and design the kernel function of the SVM model by combining the penalty factor, scaling, polynomial order, and weight coefficient; Step 4: Adopt the Dwarf Hippopotamus algorithm, give the Hippopotamus optimization algorithm a vigilant attribute, introduce a warning factor, improve the position update method of the Hippopotamus during defense, use the samples containing feature vectors and radiation source working modes in the training set to train the SVM classification model, and optimize the penalty factor, scaling, polynomial order, and weight coefficient of the SVM classification model; Step 5: Obtain the pulse description word of the radiation source to be identified, obtain the corresponding two-dimensional image through data conversion, and obtain the corresponding feature vector through the pre-trained feature extraction model; input the feature vector into the trained SVM classification model to obtain the recognition result of the radiation source working mode.

2. A radiation source working mode recognition method according to claim 1, characterized in that: The step 4 is specifically as follows: Step 4.1: Initialize the number of iterations t = 1, set the maximum number of iterations T_Maxiter, and the population size N of the pygmy hippopotamus; each pygmy hippopotamus individual is composed of the penalty factor, telescoping scale, polynomial order, and weight coefficient of the SVM model; initialize the positions of N pygmy hippopotamus individuals, and calculate the fitness Fit corresponding to the position of each pygmy hippopotamus individual z , take the fitness Fit z The position of the largest Pygmy Hippopotamus individual is taken as the initial global optimal position Dhippo(0); Initialize the positions of N Pygmy Hippopotamus individuals as: Hippo z (1)=[q z1 ,q z2 ,q z3 ,q z4 ] what zm =l m +η1(u m -l m ) Among them, q z1 ,q z2 ,q z3 ,q z4 They represent the four parameters of the zth pygmy hippopotamus, l m Indicates the lower bound of the corresponding parameter, u m The upper bound of the corresponding parameter value is m=1,2,3,4; η1 is a random number in [0,1]; The fitness z To: Set the position of the zth pygmy hippopotamus individual Hippo z The corresponding penalty factor, scaling scale, polynomial order and weight coefficient are substituted into the SVM model, and the recognition accuracy of the SVM model is obtained through the training set Fit z ; Step 4.2: Enter the exploration phase, record the first N / 2 pygmy hippopotamuses as males and the last N / 2 pygmy hippopotamuses as females, and update the position of each pygmy hippopotamus in the exploration phase according to gender; in, is the position of the z1th male pygmy hippopotamus in the exploration phase of the tth iteration; is the position of the z2th female pygmy hippopotamus in the exploration phase of the tth iteration; G is a random number in [0,1]; Q is a random number that follows a normal distribution, η2 is a random number in [0,1], η3 is a random number in (0,1], and rand(1,2) is a random number in [1,2]; h1 is an environmental factor, h1=η4+Q, η4 is a random number in [0,1]; T is a distance factor, Step 4.3: Calculate the fitness of each male and female pygmy hippopotamus according to their positions during the exploration phase. If the maximum fitness among them is greater than the corresponding fitness of the global optimal position Dhippo(t-1), then update the global optimal position. Step 4.4: Enter the predation stage and generate predators that can threaten the survival of the pygmy hippopotamus. The position of the predator in the search space can be expressed as: Predator(t)=[p t1 ,p t2 ,p t3 ,p t4 ] p tm =l m +η5(u m -l m ) Among them, η5 is a random number in [0,1]; Step 4.5: Calculate the fitness of the predator's position Predator(t), according to the warning factor vi of each pygmy hippopotamus at the previous moment z (t-1) Update the position of each pygmy hippopotamus in the population; Where RL is a random variable with Levy distribution, η6 is a uniform random number in [2,4], η7 is a uniform random number in [2,3], η8 is a uniform random number in [-1,1], η9 is a random number in [0,1], D is the distance between the predator's position and the global optimal position, D = || Predator(t) - Dhippo(t-1)||; Fit Predator(t) <Fit Dhippo(t-1) It means that the fitness corresponding to the predator's position Predator(t) is less than the fitness corresponding to the global optimal position Dhippo(t-1); Step 4.6: Update the early warning factors of each pygmy hippopotamus; Step 4.7: Update the position of each hippopotamus according to the Hippo z (t+1) Calculate fitness and update the global optimal position; Step 4.8: If t<T_Maxiter, set t=t+1 and return to step 4.2 for the next iteration; otherwise, output the global optimal position and use the penalty factor, scaling scale, polynomial order and weight coefficient of the corresponding SVM model to construct the optimal SVM classification model.

3. A radiation source working mode recognition method according to claim 1, characterized in that: In step 3, the kernel function of the SVM model designed by combining the penalty factor C, the telescopic scale σ, the polynomial order d and the weight coefficient λ is specifically: The optimal margin hyperplane y(x) of the SVM model is expressed as: y(x)=sgn(ω·δ(x)+b) Where ω is the normal vector of the hyperplane; b is the offset of the hyperplane; sgn(·) is the step function; δ(·) is the feature mapping function of the input sample; x is the vector form of the input sample; y is the vector form of the sample category; The objective function of SVM model optimization is expressed as: Among them, ξ j is the relaxation factor, ξ j ≥0, j=1,2,...,n, n is the sample length; The constraints for SVM model optimization are expressed as: y j (oh T ·δ(x j )+b)≥1-ξ j In the formula, x j is the jth element in x; y j is the jth element in y; Construct the kernel function of the SVM model: In the formula, x T is the transpose of x; ||·|| is the modulo operation; exp(·) is the exponential function; According to the constructed SVM model kernel function, the normal vector ω and offset b of the hyperplane are expressed as: Among them, a j is the first dual variable to be solved, a=[a1,a2,...,a n ] T is the set of solutions that satisfy the constraints.

4. The method for identifying a radiation source working mode according to claim 1, characterized in that: In step 1, a training set consisting of samples including pulse description words and radiation source working modes is obtained, specifically: Generates pulse description words PDW containing various types of radiation source working modes: PDW=[PDW1,PDW2,...,PDW i ,...,PDW num ] Among them, PDW i is the pulse description word of the ith pulse, PDW i =[PW i ,BW i ,CF i ,PRI i ], PW i is the pulse width of the ith pulse; BW i is the pulse bandwidth of the ith pulse; CF i is the pulse carrier frequency of the ith pulse; PRI i is the pulse repetition interval of the i-th pulse; i = 1, 2, ..., num, num is the number of pulses; Perform data mapping on all parameters in PDW to form the radiation source working mode parameter data set R_data: Among them, PW_n i is the i-th pulse width PW in PDW i Normalized results, BW_n i is the bandwidth BW of the i-th pulse in PDW i Normalized results, CF_n i is the carrier frequency CF of the i-th pulse in PDW i Normalized results, PRI_n i is the ith pulse repetition interval PRI in PDW i Normalized results, min(·) is the minimum value in the parameter set, and max(·) is the maximum value in the parameter set.

5. A method for identifying a radiation source working mode according to claim 4, characterized in that: In step 1, data conversion is performed on the samples in the training set to generate a two-dimensional image, specifically: Assume that the image sample consists of four-channel data of A, R, G, and B. The image sample data set Fig_data is represented as: Fig_data=[Fig_data1,Fig_data2,...,Fig_data fp ,...,Fig_data num ] Among them, Fig_data fp is the fpth image sample, Fig_data fp =cat(data_A,data_R,data_G,data_B), cat(·) is a channel merging function, data_A is the A channel data of the image sample, data_R is the R channel data of the image sample, data_G is the G channel data of the image sample, and data_B is the B channel data of the image sample; Where k is the number of times each column of R_data is replicated.

6. A method for identifying a radiation source working mode according to claim 1, characterized in that: The pre-trained feature extraction model in step 2 is specifically: migrating the PlainMamba network model pre-trained based on the ImageNet image dataset, removing the softmax classification layer in the PlainMamba network model, forming a migration network model model1, and taking the two-dimensional image sample as the input of the model model1.

7. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images