A method for identifying the working mode of a radiation source based on transfer learning

Abstract

The application discloses a radiation source working mode recognition method based on transfer learning, which comprises the following steps: acquiring small sample radiation source data, and selecting auxiliary data from existing radiation source working mode data; forming a source domain of transfer learning by combining the small sample radiation source data and the selected auxiliary data; taking to-be-recognized radiation source data as a target domain of transfer learning; performing manifold feature learning on samples of the source domain and the target domain; acquiring overall distribution differences of the source domain after the manifold feature learning and the target domain after the manifold feature learning, and obtaining training data; optimizing a classifier based on the training data, and acquiring a radiation source working mode of the target domain. The method solves the problem that small samples are difficult to recognize radiation source working modes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of emission source operation mode recognition, and particularly relates to a radiation source operation mode recognition method based on transfer learning. BACKGROUND

[0002] Judging and identifying the operation mode of the radiation source is an important content of electronic reconnaissance, and electronic reconnaissance is the premise and basis of electronic countermeasure. Only by correctly understanding the operation mode of the radiation source, and then inferring the function of the radiation source, evaluating its threat degree and intention, can specific measures be taken to counter the electromagnetic target of the radiation source to obtain the battlefield electromagnetic dominance. However, with the rapid development of electronic technology and the continuous change of tactics in modern war, the environment faced by electronic reconnaissance is becoming more and more complex. The intercepted signals and data of reconnaissance present the characteristics of incompleteness, insufficiency and small sample. Even if a large amount of radiation source data is obtained through reconnaissance, the sample data with labels may still present the characteristics of small sample, which causes the performance of the existing radiation source operation mode recognition method trained by a large amount of samples to sharply decrease. SUMMARY

[0003] In view of the above problems in the prior art, the radiation source operation mode recognition method based on transfer learning provided by the present application solves the problem that small samples are difficult to identify the operation mode of the radiation source.

[0004] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows:

[0005] The present application provides a radiation source operation mode recognition method based on transfer learning, which comprises the following steps:

[0006] S1, obtaining small sample radiation source data, and selecting auxiliary data from existing radiation source operation mode data;

[0007] S2, constructing a source domain for transfer learning by using the small sample radiation source data and the selected auxiliary data; and taking the radiation source data to be identified as a target domain for transfer learning;

[0008] S3, performing manifold feature learning on the samples of the source domain and the target domain to obtain a manifold feature learning source domain and a manifold feature learning target domain;

[0009] S4, obtaining the overall distribution difference of the manifold feature learning source domain and the manifold feature learning target domain;

[0010] S5, optimizing a classifier based on the overall distribution difference, and obtaining the operation mode of the radiation source of the target domain.

[0011] Further, the specific method for selecting auxiliary data from the existing radiation source operation mode data in step S1 comprises the following sub-steps:

[0012] S1-1, according to the formula:

[0013]

[0014]

[0015] Calculate the weighted Euclidean distance dist(x, y) of the sample x in the small sample radiation source data and the sample y in the existing radiation source working mode data, and then obtain the weighted Euclidean distance of a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data; Where n represents the feature dimension number of the sample; x i and y i are the i-th dimensional features of the sample x and the sample y respectively; w i is the weight of the i-th dimensional feature; W is the sum of the weights of the i-th dimensional features; e is a natural constant; and σ is a constant;

[0016] S1-2, judge whether the weighted Euclidean distance of a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than the distance threshold value, if yes, discard the sample in the existing radiation source working mode data; Otherwise, keep the sample in the existing radiation source working mode data; Enter step S1-3;

[0017] S1-3, obtain the cosine similarity of the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data by using the cosine similarity measurement method;

[0018] S1-4, judge whether the cosine similarity of the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than the similarity threshold value, if yes, discard the sample in the existing radiation source working mode data; Otherwise, keep the sample in the existing radiation source working mode data; Complete the auxiliary data selection.

[0019] Further, the specific method of manifold feature learning of the samples in the source domain and the target domain in step S3 includes the following sub-steps:

[0020] S3-1, map the samples in the source domain and the target domain to two poles of the Grassmann manifold space respectively;

[0021] S3-2, according to the formula:

[0022] <Z k , Z j >=∫0 1 (φ(t) T x k ) T (φ(t) T x j )dt

[0023] Get any two samples x in the source domain k and x j The inner product of transforming the source domain pole φ(0) to the target domain pole φ(1). <Z k Z j >;where φ(t)∈G(d), G(d) represents a d-dimensional subspace of the manifold space G, t∈[0,1];(.) T Represents the transpose of a matrix;

[0024] S3-3, According to the formula:

[0025]

[0026] Based on the kernel trick, by arbitrarily pairing two samples x in the source domain... k and x j inner product <Z k Z j > Obtain the kernel function by converting to the original space

[0027] S3-4, Kernel function Multiply by the samples in the source domain and the target domain respectively to obtain the source domain samples and the target domain samples after manifold feature learning.

[0028] Furthermore, the specific method of step S3-1 includes the following sub-steps:

[0029] S3-1-1. Merge the source domain S and the target domain T to obtain the merged domain S+T;

[0030] S3-1-2. Perform PCA dimensionality reduction on the source domain S, target domain T, and merged domain S+T to obtain the dimensionality-reduced source domain P. S The target domain P after dimensionality reduction T and the merged domain P after dimensionality reduction S+T The dimension of the reduced subspace is d.

[0031] S3-1-3. Choose the sine of the included angle as the distance and calculate P. S and P S+T The angle α between spaces d Calculate P T P S+T The angle β of space d ;

[0032] S3-1-4, According to the formula:

[0033] D(d)=0.5(sinα d +sinβ d )

[0034] the total metric D(d) of the included angle a d and the included angle b d ;

[0035] S3-1-5, select the d value that minimizes the total metric D(d), and map P S and P T corresponding to the d value as two points in the Grassmann manifold space respectively, and complete the mapping.

[0036] Further, the specific method of step S4 includes the following sub-steps:

[0037] S4-1, according to the formula:

[0038]

[0039] obtain the multi-kernel MMD mapping function φ; wherein k u is the u-th Gaussian kernel function; b u is the weight of the u-th Gaussian kernel function; and m is the total number of Gaussian kernel functions used;

[0040] S4-2, according to the formula:

[0041]

[0042]

[0043]

[0044]

[0045] d M = 2(1-2e(h))

[0046] d c = 2(1-2e(h) (c) )

[0047] obtain the total distribution difference of the source domain D s after manifold feature learning and the target domain D t after manifold feature learning wherein D f (D s , D t ) represents the edge distribution difference of D s and D t ; D f (C) (D s , D t ) represents the edge distribution difference of D s and D tthe conditional distribution difference; C is the total number of categories; d M and d c are intermediate parameters; ε(h) is the error of distinguishing the source domain after manifold feature learning and the target domain after manifold feature learning; ε(h) (c) represents the error of D s under category c and D t under category c. represents the jth sample in the source domain D s after manifold feature learning. represents the jth sample in the target domain D t after manifold feature learning; n s represents the total number of samples in D s ; n t represents the total number of samples in D t ; φ(.) represents a multi-kernel MMD mapping function. represents the jth sample in D s under category c. represents the jth sample in D t under category c. is the total number of samples in D s under category c. is the total number of samples in D t under category c; μ is an edge distribution difference and conditional distribution difference weight adjustment coefficient; ||.|| represents a norm.

[0048] Further, the specific method of step S5 includes the following sub-steps:

[0049] S5-1, according to the formula:

[0050]

[0051]

[0052]

[0053]

[0054] constructing a total loss function L of the classifier; wherein argmin represents a target function; y j represents the label of sample ; J is the total number of samples in the source domain; is the predicted label of sample by the classifier; η is a model regularization weight; ||.|| represents a norm; f represents the classifier; λ is a distribution difference weight; Y represents a (n s +n t )*1-dimensional label matrix, the first n sThe dimension is the label of the source domain sample, and the last n is the label of the source domain sample. t The dimension is 0; β is the coefficient matrix; (.) T Denotes the transpose of the matrix; K is the Gaussian kernel matrix; A represents (n s +n t )*(n s +n t A 3D diagonal matrix, where all off-diagonal elements are 0, and the first n elements of the diagonal are... s The value of one element is 1, and the value of the rest is 0; tr(.) represents the trace of the matrix; M is the maximum mean difference matrix; z k and z j Let M0 represent two samples after manifold feature learning; kj and (M) C ) kj They are matrices M0 and M respectively. c The element in the k-th row and j-th column; and They represent D under category c respectively. s The overall sample and D t The overall sample;

[0055] S5-2, According to the formula:

[0056]

[0057] Take the derivative of the overall loss function L of the classifier, obtain the value of the overall loss function L when the derivative is 0, and take the coefficient matrix β corresponding to the minimum value of the overall loss function L. * ;

[0058] S5-3, According to the formula:

[0059]

[0060] The input to the classifier is the sample z in the target domain after manifold feature learning. input At that time, the corresponding radiation source operating mode prediction result f(z) input That is, to obtain the sample z input The corresponding radiation source operating mode prediction results for the corresponding samples in the original target domain; whereby This indicates that the Gaussian kernel matrix is ​​in the input z. input and The output at that time.

[0061] The beneficial effects of this invention are: under different small sample sizes, this method has good robustness, and its recognition accuracy is basically stable at around 83%, thus enabling efficient identification of radiation source working modes under small sample conditions. Attached Figure Description

[0062] Figure 1 Fig. 1 is a flowchart of the method;

[0063] Fig. 2 is a visualization of the radiation source data in the embodiment; wherein Fig. 2(a) is a visualization of the target domain features, Fig. 2(b) is the source domain under bias amount 0.3, and Fig. 2(c) is the source domain under bias amount 0.5;

[0064] Figure 3 Fig. 3 is the classification model accuracy under the condition of 10-100 small samples when the bias amount intensity is 0.3.

[0065] Figure 4 Fig. 4 is the classification model accuracy under the condition of 10-100 small samples when the bias amount intensity is 0.5. DETAILED DESCRIPTION

[0066] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0067] As shown in the figure, the radiation source working mode recognition method based on transfer learning includes the following steps: Figure 1

[0068] S1, obtaining small sample radiation source data, and selecting auxiliary data from existing radiation source working mode data;

[0069] S2, constructing the source domain of transfer learning by the small sample radiation source data and the selected auxiliary data; and taking the to-be-recognized radiation source data as the target domain of transfer learning;

[0070] S3, performing manifold feature learning on the samples of the source domain and the target domain to obtain the source domain after manifold feature learning and the target domain after manifold feature learning;

[0071] S4, obtaining the overall distribution difference of the source domain after manifold feature learning and the target domain after manifold feature learning;

[0072] S5, optimizing the classifier based on the overall distribution difference, and obtaining the radiation source working mode of the target domain.

[0073] The specific method of selecting auxiliary data from the existing radiation source working mode data in step S1 includes the following sub-steps:

[0074] S1-1, according to the formula:

[0075] ​

[0076]

[0077] calculating a weighted Euclidean distance dist(x, y) of the sample x in the small sample radiation source data and the sample y in the existing radiation source working mode data, and then obtaining the weighted Euclidean distance of a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data; wherein n represents the characteristic dimension number of the sample; x i and y i are the i-th dimensional characteristics of the sample x and the sample y respectively; w i is the weight of the i-th dimensional characteristic; W is the total weight of the i-th dimensional characteristic; e is a natural constant; σ is a constant; and specifically, the value can be 0.02;

[0078] S1-2, judging whether the weighted Euclidean distance of a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than a distance threshold value, if yes, discarding the sample in the existing radiation source working mode data; otherwise, retaining the sample in the existing radiation source working mode data; and entering step S1-3;

[0079] S1-3, obtaining the cosine similarity of the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data by using a cosine similarity measurement method;

[0080] S1-4, judging whether the cosine similarity of the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than a similarity threshold value, if yes, discarding the sample in the existing radiation source working mode data; otherwise, retaining the sample in the existing radiation source working mode data; and completing the auxiliary data selection.

[0081] The specific method for manifold feature learning of the samples in the source domain and the target domain in step S3 includes the following sub-steps:

[0082] S3-1, mapping the samples in the source domain and the target domain to two poles of the Grassmann manifold space respectively;

[0083] S3-2, according to the formula:

[0084] <Z k , Z j >=∫0 1 (φ(t) T x k ) T (φ(t) T x j )dt

[0085] obtaining any two samples x k and xj The inner product of transforming the source domain pole φ(0) to the target domain pole φ(1). <Z k Z j >;where φ(t)∈G(d), G(d) represents a d-dimensional subspace of the manifold space G, t∈[0,1];(.) T Represents the transpose of a matrix;

[0086] S3-3, According to the formula:

[0087]

[0088] Based on the kernel trick, by arbitrarily pairing two samples x in the source domain... k and x j inner product <Z k Z j > Obtain the kernel function by converting to the original space

[0089] S3-4, Kernel function Multiply by the samples in the source domain and the target domain respectively to obtain the source domain samples and the target domain samples after manifold feature learning.

[0090] The specific method of step S3-1 includes the following sub-steps:

[0091] S3-1-1. Merge the source domain S and the target domain T to obtain the merged domain S+T;

[0092] S3-1-2. Perform PCA dimensionality reduction on the source domain S, target domain T, and merged domain S+T to obtain the dimensionality-reduced source domain P. S The target domain P after dimensionality reduction T and the merged domain P after dimensionality reduction S+T The dimension of the reduced subspace is d.

[0093] S3-1-3. Choose the sine of the included angle as the distance and calculate P. S and P S+T The angle α between spaces d Calculate P T P S+T The angle β of space d ;

[0094] S3-1-4, According to the formula:

[0095] D(d)=0.5(sinα d +sinβ d )

[0096] Obtain the included angle α d and the included angle β dtotal metric D(d) of Ds and D

[0097] S3-1-5, selecting the d value that minimizes the total metric D(d), and selecting P S and P T respectively as two points in the Grassmann manifold space respectively mapped by the samples in the source domain and the target domain, and completing the mapping.

[0098] The specific method of step S4 includes the following sub-steps:

[0099] S4-1, according to the formula:

[0100]

[0101] Obtain the multi-kernel MMD mapping function φ; wherein k u is the u-th Gaussian kernel function; β u is the weight of the u-th Gaussian kernel function; and m is the total number of Gaussian kernel functions used;

[0102] S4-2, according to the formula:

[0103]

[0104]

[0105]

[0106]

[0107] d M = 2(1-2ε(h))

[0108] d c = 2(1-2ε(h) (c) )

[0109] Obtain the source domain Ds after manifold feature learning and the target domain D t after manifold feature learning wherein D f (D s , D t ) represents the edge distribution difference of D s and D t ; D f (C) (D s , D t ) represents the conditional distribution difference of D s and D t ; C is the total number of categories; d M and d care intermediate parameters; ε(h) is the error of distinguishing the source domain after manifold feature learning and the target domain after manifold feature learning; ε(h) (c) represents the jth sample in D s under the category c; t error of D represents the jth sample in D s after manifold feature learning; represents the jth sample in D t after manifold feature learning; n s represents the total number of samples in D s ; n t represents the total number of samples in D t ; φ(.) represents a multi-kernel MMD mapping function; represents the jth sample in D s under the category c; represents the jth sample in D t under the category c; is the total number of samples in D s under the category c; is the total number of samples in D t under the category c; μ is an edge distribution difference and conditional distribution difference weight adjustment coefficient; ||.|| represents a norm.

[0110] The specific method of step S5 includes the following sub-steps:

[0111] S5-1, according to the formula:

[0112]

[0113]

[0114]

[0115]

[0116] constructing a total loss function L of the classifier; wherein argmin represents a target function; y j represents the label of sample ; J is the total number of samples in the source domain; is the predicted label of sample by the classifier; η is a model regularization weight; ||.|| represents a norm; f represents the classifier; λ is a distribution difference weight; Y represents a (n s +n t )*1-dimensional label matrix, the first n s dimensions are the labels of the source domain samples, and the last n t dimensions are 0; β is a coefficient matrix; (.) TDenotes the transpose of the matrix; K is the Gaussian kernel matrix; A represents (n s +n t )*(n s +n t A 3D diagonal matrix, where all off-diagonal elements are 0, and the first n elements of the diagonal are... s The value of one element is 1, and the value of the rest is 0; tr(.) represents the trace of the matrix; M is the maximum mean difference matrix; z k and z j Let M0 represent two samples after manifold feature learning; kj and (M) C ) kj They are matrices M0 and M respectively. c The element in the k-th row and j-th column; and They represent D under category c respectively. s The overall sample and D t The overall sample;

[0117] S5-2, According to the formula:

[0118]

[0119] Take the derivative of the overall loss function L of the classifier, obtain the value of the overall loss function L when the derivative is 0, and take the coefficient matrix β corresponding to the minimum value of the overall loss function L. * ;

[0120] S5-3, According to the formula:

[0121]

[0122] The input to the classifier is the sample z in the target domain after manifold feature learning. input At that time, the corresponding radiation source operating mode prediction result f(z) input That is, to obtain the sample z input The corresponding radiation source operating mode prediction results for the corresponding samples in the original target domain; whereby This indicates that the Gaussian kernel matrix is ​​in the input z. input and Output at that time.

[0123] In one embodiment of the present invention, a simulation dataset of radiation sources with four different operating modes was used for experimental analysis to verify the effectiveness and feasibility of the knowledge transfer model. The radiation source dataset contains a total of 291,720 samples, categorized into four labels: Category 1 (104,491 samples), Category 2 (77,196 samples), Category 3 (110,033 samples), and Category 4 (10,899 samples). Each sample includes four features: carrier frequency, amplitude, pulse width, and pulse repetition interval.

[0124] Experimental design: The experiment verifies the difference in classification accuracy between the traditional method and the present method under the conventional characteristics (carrier frequency, amplitude, pulse width, pulse repetition interval jointly constitute), wherein the traditional method adopts SVM.

[0125] Source domain and target domain construction: Both are simulation data sets under ideal conditions, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100 samples in each class of the simulation data set are selected as known small sample radiation source data, and the small sample working mode label is known. 1000 samples in each class of radiation source are selected as 4000 samples in total as the target domain to be identified. At the same time, different intensities of bias are superimposed on the target domain to construct the source domain with similar knowledge as the target domain. The source domain construction formula is as follows:

[0126]

[0127] Wherein x ji is the i-th dimensional parameter of the j-th sample of the target domain; α is the bias intensity; represents the average value of the j-th dimensional parameter of the target domain sample; rand is a uniformly distributed 0-1 random number. is the constructed source domain.

[0128] Select the samples constructed under α = 0.3 and α = 0.5 as the screening data, and the present application uses weighted Euclidean distance to evaluate the similarity of the screening data, so as to select the samples similar to the target domain radiation source data as the standard, thereby constructing the source domain suitable for the knowledge transfer model. The radiation source data visualization is shown in Figure 2. As can be seen from Figure 2, compared with the target domain, the distribution position of each class of samples in the source domain in the feature space has changed greatly, and the classifier trained by the source domain cannot be directly used for classification of the target domain samples. Secondly, with the increasing intensity of the bias, the distinguishability between different classes of source domain samples becomes weaker, and the difficulty of constructing a classifier using source domain samples increases.

[0129] Classification result comparison: based on the given source domain and target domain, the accuracy of the target domain sample classification by the present embodiment using SVM and the present method is shown in Table 1 and Table 2 (wherein 10, 20, …, 100 are the sample quantities).

[0130] Table 1: Recognition accuracy (%) of each classification method under bias = 0.3

[0131]

[0132] Table 2: Recognition accuracy (%) of each classification method under bias = 0.5

[0133]

[0134] like Figure 3 and Figure 4 As shown, with bias strengths of 0.3 and 0.5, this method achieved the highest classification accuracy under different sample sizes. This phenomenon indicates that the method can effectively reduce the distribution difference between the source and target domains, enabling the classifier trained using source domain samples to still achieve good recognition performance in the target domain. Furthermore, under different small sample sizes, the patented model exhibits good robustness, with its recognition accuracy remaining relatively stable at around 83%.

Claims

1. A method for identifying the operating mode of a radiation source based on transfer learning, characterized in that, The method comprises the following steps: S1, acquiring small sample radiation source data, and selecting auxiliary data from existing radiation source working mode data; S2, constructing a source domain for transfer learning by combining the small sample radiation source data and the selected auxiliary data; regarding to-be-identified radiation source data as a target domain for transfer learning; S3, performing manifold feature learning on samples in the source domain and the target domain to obtain a source domain after manifold feature learning and a target domain after manifold feature learning; S4, acquiring a total distribution difference between the source domain after manifold feature learning and the target domain after manifold feature learning; S5, optimizing a classifier based on the total distribution difference, and acquiring a radiation source working mode of the target domain; The specific method for selecting auxiliary data from the existing radiation source working mode data in step S1 comprises the following sub-steps: S1-1, according to the formula: Computing the weighted Euclidean distance between a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data x And the existing radiation source working mode data sample y Weighted Euclidean distance , and then get the weighted Euclidean distance between a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data; wherein n Indicates the number of feature dimensions of the sample; And Respectively, the first i Dimensional features of sample x And sample y ​ For the first i weight of the feature; W is the sum of the weights of the features; i the sum of the weights of the features; e is a natural constant; is a constant; S1-2, judging whether the weighted Euclidean distance between a single sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than a distance threshold value, if yes, discarding the sample in the existing radiation source working mode data; otherwise, retaining the sample in the existing radiation source working mode data; and entering step S1-3; S1-3, acquiring the cosine similarity between the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data by using a cosine similarity measurement method; S1-4, judging whether the cosine similarity between the retained sample in the existing radiation source working mode data and all samples in the small sample radiation source data is greater than a similarity threshold value, if yes, discarding the sample in the existing radiation source working mode data; otherwise, retaining the sample in the existing radiation source working mode data; and completing the selection of auxiliary data; The construction expression of the sample in the existing radiation source working mode data is: wherein is the i-th dimensional parameter of the j-th sample of the target domain; is the bias strength; represents the average value of the j-th dimensional parameter of the target domain sample; and rand is a random number uniformly distributed between 0 and 1; is the sample in the existing radiation source working mode data.

2. The method of claim 1, wherein, The specific method for performing manifold feature learning on samples in the source domain and the target domain in step S3 comprises the following sub-steps: S3-1, mapping the samples in the source domain and the target domain to two poles of a Grassmann manifold space respectively; S3-2, according to the formula: Obtaining any two samples in the source domain and from the pole of the source domain Transforming to the pole of the target domain inner product ; wherein , represents a d dimensional subspace in the manifold space G, ; represents the transpose of the matrix; S3-3, according to the formula: Based on kernel tricks, by arbitrarily selecting two samples from the source domain... and inner product Obtaining kernel functions by converting to the original space ; S3-4, multiplying the kernel function respectively with the samples in the source domain and the target domain, respectively to obtain the source domain samples after manifold feature learning and the target domain samples after manifold feature learning.

3. The method of claim 2, wherein the method further comprises: The specific method of step S3-1 comprises the following sub-steps: S3-1-1, combining the source domain S and the target domain T to obtain a combined domain S+T; S3-1-2, performing PCA dimension reduction on the source domain S, the target domain T and the merged domain S+T to obtain the reduced source domain , the reduced target domain and the reduced merged domain ; wherein the dimension of the reduced subspace is d ; S3-1-3, select the sine value of the included angle as the distance, calculate and the included angle of the space , calculate the included angle of the space ; S3-1-4, according to the formula: angle of acquisition and the angle of the total measure ; S3-1-5, selecting a total metric minimum d value, the d value corresponding and respectively as the source domain and the target domain in the sample respectively mapped in the Grassmann manifold space two points, complete mapping.

4. The method of claim 2, wherein, The specific method of step S4 comprises the following sub-steps: S4-1, according to the formula: Obtaining a multi-kernel mmd mapping function ; wherein is a first u gaussian kernel function; is a weight of a first u gaussian kernel function; m is a total number of gaussian kernel functions employed; S4-2, according to the formula: Obtain the source domain after manifold feature learning and the target domain after manifold feature learning Overall distribution differences ;in express and Differences in marginal distribution; express and The conditional distribution differences; C Total number of categories; and These are all intermediate parameters; To distinguish the error between the source domain after manifold feature learning and the target domain after manifold feature learning; Indicate category c Down With category c Down The error; Represents the source domain after manifold feature learning. The first in j One sample; This represents the target domain after manifold feature learning. The first in j One sample; express Total number of samples; express Total number of samples; Represents a multi-core MMD mapping function; Indicate category c Down The first in j One sample; Indicate category c Down The first in j One sample; For category c Down The total number of samples in the sample; For category c Down The total number of samples in the sample; This is a weighting adjustment coefficient for the differences in marginal distribution and conditional distribution; Represents the norm.

5. The method of claim 4, wherein, The specific method of step S5 comprises the following sub-steps: S5-1, according to the formula: Construct the overall loss function of the classifier L ;in Represent the objective function; Indicates sample The tag, J This represents the total number of samples in the source domain. For the classifier to classify samples Predicted labels; Use these weights to regularize the model; Represents the norm; f Represents a classifier; Weights for distribution differences; Y express A dimensional label matrix, front The dimension represents the label of the source domain sample, followed by... Dimension is 0; It is a coefficient matrix; Represents the transpose of a matrix; It is a Gaussian kernel matrix; express A 3D diagonal matrix, where all off-diagonal elements are 0, and the first element of the diagonal is 0. One element has a value of 1, and the rest have a value of 0; Represents the trace of a matrix; This is the matrix of maximum mean difference; and These represent two samples after manifold feature learning; and Each is a matrix and The Middle k Line 1 j Column elements; and Representing categories c Down The overall sample and The overall sample; S5-2, according to the formula: Overall loss function for the classifier L Derivation, get the value of the overall loss function when the derivative is 0 L Take the minimum value of the overall loss function L Corresponding coefficient matrix when the overall loss function ; S5-3, according to the formula: The input to the classifier is the samples in the target domain after manifold feature learning. At that time, the corresponding radiation source operating mode prediction results That is, to obtain the sample The corresponding radiation source operating mode prediction results for the corresponding samples in the original target domain; whereby Indicates the Gaussian kernel matrix at the input and The output at that time.

Images