Open environment radiation source individual identification method based on manifold learning

Abstract

The invention discloses an open environment radiation source individual identification method based on manifold learning, and belongs to the technical field of radiation source individual identification. According to the method, firstly, high-dimensional signal data are mapped to a low-dimensional Grassmann manifold space, and the internal structure of the data is kept while redundant features in the high-dimensional data are removed; and then aligning the open manifold structures of the source domain and the target domain in the low-dimensional manifold space, so that the method can measure the distance between the prototypes of the source domain and the target domain in the unified manifold space. And finally, performing cross-domain prototype matching by replacing Euclidean distance with geodesic distance on a Grassmann manifold, labeling a pseudo-label for target domain data according to a matching result, and jointly using source domain label data and target domain pseudo-label data as a reference set of a nearest neighbor classifier so as to obtain a nearest neighbor classifier. And accurate identification of known and unknown types of radiation source individuals in an open environment is realized through a nearest neighbor classification strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radiation source individual identification, and particularly relates to an open environment radiation source individual identification method based on manifold learning. BACKGROUND

[0002] Specific Emitter Identification (SEI) is a technology that uses the difference of radio frequency fingerprinting (RFF) of wireless signals to distinguish different radiation source individuals, and has important application value in military and civilian fields. Existing SEI researches mostly focus on designing better radio frequency fingerprint feature extraction and classification methods to achieve accurate identification of radiation source individuals. These researches often assume that the SEI model is always deployed in an ideal closed environment, that is, the radiation source individuals to be identified by the model in the training stage and the test stage come from the same category set, and the data used by the model in the training stage and the test stage come from the same channel environment. However, the above assumption is not true in an open environment.

[0003] On the one hand, due to the rapid evolution of wireless communication technology and the explosive growth of Internet of Things devices, in an open environment, new radiation source individuals may be introduced due to device technology updates, device type expansion or dynamic deployment of devices. Most of the existing research work is based on closed set classification assumption, that is, the data used by the model in the training stage and the test stage come from the same known category set. However, since the radiation source individuals in the open environment are always changing, no matter how large the known category set corresponding to the training data is, the model may encounter unknown category data in the test stage, at this time, since the model only has closed set classification ability, it can only incorrectly classify the unknown radiation source individuals into known categories, ultimately leading to a decrease in the credibility of the SEI model identification result.

[0004] On the other hand, the channel environment (such as weather conditions, multipath effects, noise and electromagnetic interference, etc.) experienced during signal transmission can cause signal distortion. The signal distortion caused by the channel environment will be extracted into the channel characteristics by the model and mixed into the radio frequency fingerprint characteristics. The channel characteristics reflect the characteristics of the channel environment and will change with the change of the channel environment. Therefore, for the signals emitted by the same radiation source, if the channel environment experienced by the signals during transmission is different, the radio frequency fingerprint characteristics extracted by the SEI model will have distribution differences. In an open environment, because the channel environment will change over time and space, the data used by the model in the training stage and the test stage are most likely to come from different channel environments, resulting in distribution differences in the radio frequency fingerprint characteristics extracted by the model in the two stages, which violates the assumption that the training set and the test set data are independent and identically distributed required by traditional machine learning, resulting in a decrease in the recognition accuracy of the SEI model in the cross-channel scenario.

[0005] To solve the problems faced by SEI in an open environment, the existing SEI method measures the similarity between data based on the Euclidean distance, and then uses the open set recognition (OSR) technology to distinguish the known and unknown category radiation source individuals in the test stage data according to the similarity size. At the same time, through the unsupervised domain adaptation (UDA) technology, the distance between the known category radio frequency fingerprint characteristics in the training and test stages is reduced, thereby alleviating the distribution offset of the radio frequency fingerprint characteristics caused by the change of the channel environment. However, the radio frequency signal of the radiation source has an inherent low-dimensional nonlinear manifold structure, and the Euclidean distance cannot truly reflect the similarity between data on the low-dimensional nonlinear manifold, resulting in that the existing SEI method cannot accurately distinguish the known and unknown category radiation source individuals in an open environment, and cannot effectively alleviate the distribution offset of the radio frequency fingerprint characteristics caused by the change of the channel environment. Therefore, it is necessary to study a radiation source individual identification method in an open environment in view of the low-dimensional nonlinear manifold structure of the radio frequency signal of the radiation source. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides an open environment radiation source individual identification method based on manifold learning to improve the recognition accuracy of known and unknown category radiation source individuals in an open environment.

[0007] The technical scheme adopted by the present application is: an open environment radiation source individual identification method based on manifold learning, which comprises the following steps:

[0008] Step 1: Collect radiation signal data of a plurality of known radiation source individuals in a period of time based on a signal receiving device, and pre-process each radiation signal data to obtain a plurality of data samples of each radiation source individual and a radiation source individual category label of each data sample, so as to obtain a source domain data set;

[0009] In an open environment, collect radiation signal data of a plurality of radiation source individuals in a period of time based on a signal receiving device, and pre-process each radiation signal data to obtain a plurality of data samples of each radiation source individual, so as to obtain a target domain data set;

[0010] Step 2: Construct a radio frequency fingerprint feature extraction network for extracting deep radio frequency fingerprint features of the radiation signal, and train the feature extraction network using the source domain data set;

[0011] Step 3: Extract deep radio frequency fingerprint features of the source domain data set and the target domain data set using the radio frequency fingerprint feature extraction network trained in step 2, and reduce the deep radio frequency fingerprint features to low-dimensional Grassmann manifold to obtain low-dimensional manifold features of the source domain and the target domain;

[0012] Then, map the low-dimensional manifold features of the source domain to the target domain manifold, match the low-dimensional manifold feature prototypes of the source domain and the target domain based on the geodesic distance on the Grassmann manifold, and label pseudo-labels for the target domain data according to the matching result;

[0013] Step 4: Jointly use the labeled data of the source domain and the pseudo-label data of the target domain as a reference set of a nearest neighbor classifier, and predict and output the radiation source individual identification in an open environment through a nearest neighbor classification strategy.

[0014] Further, the data pre-processing includes: intercepting M seconds of radiation source signal data as a data sample, and normalizing the intercepted signal data to linearly map the data to the [0, 1] interval.

[0015] Further, in step 2, the radio frequency fingerprint feature extraction network is connected to a classifier after training, which is used to predict the radiation source individual category label, and the radio frequency fingerprint feature extraction network is trained based on the classifier.

[0016] Further, the classifier for training the radio frequency fingerprint feature extraction network adopts a single-layer fully connected layer structure.

[0017] Further, the radio frequency fingerprint feature extraction network includes a convolutional network module, a plurality of residual network modules, and an average pooling layer.

[0018] Further, in step 3, reducing the deep radio frequency fingerprint features to low-dimensional Grassmann manifold specifically includes:

[0019] calculating the autocorrelation matrix of the deep radio frequency fingerprint features of the source domain and the target domain, and performing singular value decomposition on the autocorrelation matrix;

[0020] taking the first K columns of the left singular matrix obtained by the singular value decomposition to obtain a low-dimensional feature representation of the data sample on the Grassmann manifold; wherein, is a preset value. Preferably, the value of K can be determined according to the principle that the energy proportion of the first K singular values exceeds a specified proportion (90%).

[0021] Further, step 3 further includes:

[0022] setting an optimal transport model of the source domain and the target domain on the Grassmann manifold:

[0023]

[0024] wherein, and respectively represent the source domain data and the target domain data, , respectively represent the feature index of the source domain and the target domain, represents an optimal transport plan of the source domain low-dimensional feature distribution to the target domain low-dimensional feature distribution, represents a transport pair between the source domain and the target domain, represents a mapping function of the source domain low-dimensional feature to the target domain manifold, is the mapped source domain low-dimensional feature (the source domain is mapped to the target domain), is the corresponding optimal mapped source domain low-dimensional feature, is the transport cost, represents the entropy of the transport plan of the source domain low-dimensional feature distribution to the target domain low-dimensional feature distribution is a preset weight parameter for controlling the sparsity of the transport plan ; and

[0025] based on the solving result of the optimal transport model, the low-dimensional manifold feature of the source domain is mapped to the target domain manifold to obtain the mapped source domain low-dimensional feature.

[0026] Further, the transport cost is specifically set as:

[0027]

[0028] wherein, represents the target domain low-dimensional feature of the data sample j, that is, the low-dimensional manifold feature of the target domain,​​​ denotes the mapped source domain low-dimensional feature of data sample i; is the transmission cost of source domain low-dimensional feature distribution to target domain low-dimensional feature distribution, can be understood as the maximum value of all known classes ; is the unknown class penalty coefficient, is the geodesic distance on the Grassmann manifold.

[0029] Further, the unknown class penalty coefficient is the minimum reconstruction error , denotes the mapped source domain low-dimensional feature prototype, denotes the mapped source domain low-dimensional feature prototype (referred to as source domain prototype) of data sample i, is the class radius of source domain class , the class closest to the target domain low-dimensional feature distance of the mapped source domain prototype, wherein the class radius of any class is defined as: , is the average geodesic distance of intra-class low-dimensional features to low-dimensional feature prototypes, is the standard deviation of the distance, is a coefficient for controlling the coverage range of the radius.

[0030] Further, in step 3, the low-dimensional manifold feature prototypes of the source domain and the target domain are matched based on the geodesic distance on the Grassmann manifold, and pseudo labels are labeled for the target domain data according to the matching result, specifically including:

[0031] The low-dimensional manifold features of different class samples of the source domain data set are calculated , and the low-dimensional manifold feature prototypes of the source domain are obtained ;

[0032] The target domain low-dimensional manifold feature prototypes are obtained by spectral clustering on the Grassmann manifold ;

[0033] Cross-domain prototype matching is performed based on the nearest neighbor strategy, and the matching process is as follows:

[0034] For each source domain prototype , find the target domain prototype closest to it in geodesic distance in the target domain, wherein denotes the index of the found target domain data sample;

[0035] For each target domain prototype Similarly, find the source domain prototype with the closest geodesic distance in the target domain. ,in, This represents the index of the data sample found in the source domain;

[0036] Keep the bidirectional consistent prototypes in the pairing list middle: in, This indicates a bidirectional, consistent prototype pairing;

[0037] Based on the pairing list The matching relationship in the target domain is used to label pseudo-labels for data samples of known categories; that is, the labels of the matched source domain are used as pseudo-labels. The pseudo-labels are assigned to data samples of known categories in the target domain. For unmatched target domain prototypes, the category is labeled as needed after the largest known category label number (i.e., the corresponding category label number is set after the largest known category label number), and the pseudo-labels are assigned to the corresponding data samples based on the currently labeled category.

[0038] The technical solution provided by this invention brings at least the following beneficial effects:

[0039] This invention adapts to the low-dimensional nonlinear manifold structure characteristics of radio frequency signals from radiation sources and overcomes the problems of unknown category identification and radio frequency fingerprint feature distribution offset in open environments. The SEI model trained by this method can accurately identify known and unknown category radiation source individuals in open environments. First, this invention maps high-dimensional signal data to a low-dimensional Grassmann manifold space, removing redundant features from the high-dimensional data while preserving its inherent structure. Then, it aligns the open set manifold structures of the source and target domains within the low-dimensional manifold space, enabling the method to measure the distance between the source and target domain prototypes within a unified manifold space. Finally, it performs cross-domain prototype matching by replacing Euclidean distance with geodesic distance on the Grassmann manifold. Based on the matching results, pseudo-labels are assigned to the target domain data. The source domain labeled data and the target domain pseudo-label data are used together as the reference set for the nearest neighbor classifier. Through the nearest neighbor classification strategy, accurate identification of known and unknown category radiation source individuals in open environments is achieved. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1A schematic diagram of an open environment radiation source individual identification method based on manifold learning involved in the present application.

[0042] Figure 2 A radio frequency fingerprint feature extraction network structure diagram involved in the present application. DETAILED DESCRIPTION

[0043] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described in detail and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Generally, the components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application.

[0044] The embodiments of the present application provide an open environment radiation source individual identification method based on manifold learning. In one embodiment, referring to Figure 1 The implementation steps include:

[0045] Step 1: Based on the signal receiving device, collect radiation signal data of a plurality of known radiation source individuals in a period of time, and perform data preprocessing on each collected radiation signal data to obtain a plurality of data samples of each radiation source individual and a radiation source individual category label of each data sample. Then, all the labeled data samples are divided into a training set, a validation set and a test set according to a specified ratio (such as 8:1:1) to form a source domain training data set. In an actual open environment, based on the signal receiving device, collect radiation signal data of a plurality of radiation source individuals in a period of time, and perform data preprocessing on each collected radiation signal data to obtain a plurality of data samples of each radiation source individual. Then, all the unlabeled data samples are divided into a training set and a test set according to a ratio of 8:2 to form a target domain test data set.

[0046] Step 2: Build a radio frequency fingerprint feature extraction network. In the present embodiment, the network structure adopts a one-dimensional ResNet-18 structure (other backbone networks can also be used based on actual needs), and the source domain data set is used to train the feature extraction network.

[0047] Step 3: Use the feature extraction network trained in step 2 to extract the deep radio frequency fingerprint features of the source domain and target domain radiation source signals. Then, the deep features are reduced to low-dimensional Grassmann manifolds. The low-dimensional manifold features of the source domain are mapped to the target domain manifold. Finally, the source domain and target domain low-dimensional feature prototypes are matched based on the geodesic distance on the Grassmann manifold, and pseudo-labels are labeled for the target domain data according to the matching results.

[0048] Step 4: Train the nearest neighbor classifier using both labeled source domain data and pseudo-labeled target domain data. Finally, use the prediction output of the nearest neighbor classifier to identify radiation source individuals in an open environment.

[0049] In one embodiment, the data preprocessing procedure in step 1 is as follows: M seconds of radiation source signal data are extracted as a data sample (M is set according to the actual characteristics of the data). In this embodiment, M is set to 3025. The extracted signal data... The minimum-maximum normalization method is used to linearly map the data to the [0,1] interval, as shown in the following formula:

[0050]

[0051] in, for The first in Data sample points, Represents the normalized i-th Each data sample point is analyzed. Based on the parameter settings of the signal receiving equipment, the system determines which category of known radiation source the data sample belongs to and assigns a true classification label to the data sample.

[0052] In one embodiment, the radio frequency fingerprint feature extraction network used in this invention is used to extract high-dimensional radio frequency fingerprint features from the radio frequency signal of the radiation source. Its structure employs a deep residual network structure, such as... Figure 2 As shown, it includes a cascaded convolutional network module and several residual network modules, and finally outputs the extracted features through an average pooling layer. Figure 2 The fully connected layer (FC) in the network is used to train the radio frequency fingerprint feature extraction network, which is used to predict the individual radiation source category of the source domain data. The convolutional network module includes cascaded convolutional layers, batch normalization (BN) layers, ReLU activation layers, and max pooling layers. The convolutional kernel of the convolutional layer can be set to 7×7, with a stride of 2 and padding of 3, having 1 input channel and 64 output channels. The max pooling layer kernel can be set to 3×3, with a stride of 2 and padding of 3. In this embodiment, the residual network module includes cascaded convolutional layers, batch normalization layers, and ReLU activation layers. The input of the convolutional layer and the output of the ReLU activation layer in the residual network module are set as short connections, meaning that the output of each residual network module is obtained based on the fusion result of the input of the convolutional layer and the output of the ReLU activation layer. In this embodiment, as shown... Figure 2As shown, the main branch of the residual network module consists of two cascaded convolutional blocks (convolutional layer, batch normalization layer, and ReLU activation layer). The convolutional kernels are all set to 3×3, the number of output channels is 64, and the stride and padding are both 1. In the short connection branch, the short connection between the input and output of the main branch is achieved through a convolutional layer with a 1×1 kernel (followed by a batch normalization layer).

[0053] Preferably, the training method for the feature extraction network in step 2 is as follows: set the learning rate to... The batch size is 64, and the total number of training iterations is 50. The optimizer used is Adam, with a weight decay factor of 1*. This is used for regularization to prevent overfitting. The classifier uses a single fully connected layer structure, with the number of classifier heads equal to the total number of classes in the training set. The cross-entropy loss function is used. During training, the learning rate is dynamically adjusted, decaying to 0.1% every 10 epochs using the StepLR learning rate scheduler.

[0054] In one embodiment, step 3 includes the following steps:

[0055] Step 3-1: Map the deep fingerprint features extracted by the radio frequency fingerprint feature extraction network to a low-dimensional Grassmann manifold.

[0056] The source domain deep fingerprint features are denoted as The target domain deep fingerprint features are (in , (representing the number of samples in the source and target domains), "", "These are used to identify the source and target domains, respectively. First, the autocorrelation matrix of the deep fingerprint features is calculated:"

[0057]

[0058] Next, the autocorrelation matrix Perform singular value decomposition:

[0059]

[0060] in, It is a left singular matrix, and its column vectors are an orthonormal basis ( ). Let the singular value matrix be the matrix. Take the left singular matrix. The former This yields a low-dimensional feature representation of the data samples on the Grassmann manifold. ,in The value of can be chosen to make the previous The principle is to determine that the energy percentage of a singular value exceeds 90%.

[0061] Step 3-2: Map the low-dimensional features of the source domain to the target domain manifold.

[0062] The low-dimensional feature representation of the source domain after mapping is obtained by solving the following optimal transport problem on the Grassmann manifold:

[0063]

[0064] in, and These represent the source and target domain data, respectively. , These represent the feature indices of the source and target domains, respectively. This represents a transport pair between the source and destination domains. Let f(x) represent the mapping function from low-dimensional features of the source domain to the manifold of the target domain. The low-dimensional features of the source domain after mapping These are the low-dimensional features of the source domain after the corresponding optimal mapping.

[0065] Because of direct solution It is quite difficult; the embodiments of the present invention use weighted averages. The mean is calculated using the following steps. First, logarithmic mappings on the Grassmann manifold are used to transform the low-dimensional features of each target domain. Mapping to tangent space:

[0066]

[0067] in, The actual calculation formula is: ,in ,in This represents singular value decomposition.

[0068] Then calculate the weighted average tangent vector:

[0069]

[0070] Finally, the tangent vector is mapped back to the manifold using the exponential mapping on the Grassmann manifold:

[0071]

[0072] in, .

[0073] This is a transport plan for the low-dimensional feature distribution of the source domain to the low-dimensional feature distribution of the target domain. This corresponds to the optimal transmission plan. The entropy regularization formula is in the form of:

[0074]

[0075] in, These are weighting parameters used to control... The sparsity, an empirical value, is taken in this embodiment as... .

[0076] The transmission cost from the low-dimensional feature distribution of the source domain to the low-dimensional feature distribution of the target domain is expressed in the form of:

[0077]

[0078] in, For unknown categories, the penalty coefficient is... The geodesic distance on the Grassmann manifold is calculated as follows:

[0079]

[0080] in, yes and The principal angles on the Grassmann manifold are obtained by calculating singular value decompositions. ,get .

[0081] This invention's embodiments calculate low-dimensional features of the target domain. Mapped source domain low-dimensional feature prototype The minimum reconstruction error between the two classes is used to determine whether a class is unknown. The minimum reconstruction error is calculated as follows:

[0082]

[0083] in, As the prototype of the mapped low-dimensional features of the source domain, the low-dimensional features of the source domain for different categories are calculated on the Grassmann manifold every 10 training epochs. The mean value is obtained by calculating the following method:

[0084]

[0085] in, For the source domain The total number of samples in each class. The method for determining an unknown class is as follows:

[0086]

[0087] in, Is with The category of the source domain prototype after the nearest mapping. For source domain category Category radius, category The radius is defined as:

[0088]

[0089] in, The average geodesic distance from the low-dimensional features within the class to the prototype. The standard deviation of the distance. The coefficient used to control the radius coverage is an empirical value; in this embodiment, it is taken as... This ensures that the radius covers 95% of the normally distributed samples. Based on this, the unknown class penalty coefficient... Defined as: .

[0090] To obtain the optimal low-dimensional features of the source domain after mapping This embodiment uses iterative optimization. and The optimal transport problem can be solved using the formula in the form of .

[0091] The iterative optimization process is as follows: First, fix... The Sinkhorn-Knopp algorithm is used to solve the optimal transport problem formula to update the algorithm. Then fix it. Calculated using the following formula Weighting in the target domain The mean has been updated :

[0092]

[0093] Repeat the above process until the loss function converges or the maximum number of iterations is exceeded, and finally obtain the low-dimensional features of the source domain mapped to the target domain manifold. .

[0094] Step 3-3: Match the low-dimensional feature prototypes of the source and target domains based on the geodesic distance on the Grassmann manifold, and label the target domain data with pseudo-labels according to the matching results.

[0095] First, the low-dimensional feature prototypes of the source and target domains are obtained separately. For the source domain prototype... Since the source domain samples are labeled, this embodiment directly calculates the low-dimensional features of samples of different categories in the source domain. The average value is obtained. For the target domain prototype, it is obtained through spectral clustering on the Grassmann manifold. The specific process is as follows: First, a symmetric similarity matrix is ​​defined. Elements in the matrix Representing low-dimensional features of the target domain and The similarity is measured by the geodesic distance on the Grassmann manifold:

[0096]

[0097] in, This is the bandwidth parameter of the Gaussian kernel, used to control the locality of local similarity. It is set according to the actual characteristics of the data; in this embodiment, it is set as follows: .

[0098] In the similarity matrix Based on this, define the degree matrix. It is a diagonal matrix, and its diagonal elements are:

[0099]

[0100] Based on this, the normalized Laplace matrix Defined as:

[0101]

[0102] After obtaining the Laplace matrix Then, calculate its preceding... The eigenvectors corresponding to the smallest eigenvalues:

[0103]

[0104] in To determine the number of target domain categories, we can iterate through the target domain clusters and select the cluster with the highest silhouette coefficient as the target domain category number. Then, we arrange the calculated feature vectors column-wise to obtain the low-dimensional embedding matrix.

[0105]

[0106] Finally, for The row vectors (i.e., the low-dimensional feature representation of the target domain samples) are clustered using K-means clustering to obtain the clustering results of the low-dimensional features of the target domain. Based on this, calculate for each cluster. The mean is used to obtain the prototype of the target domain. .

[0107] After obtaining the prototypes of the source and target domains, cross-domain prototype matching is performed based on the nearest neighbor strategy. The matching process is as follows: for each source domain prototype... Find the target domain prototype with the closest geodesic distance in the target domain. :

[0108]

[0109] For each target domain prototype Similarly, find the source domain prototype with the closest geodesic distance in the target domain. :

[0110]

[0111] Finally, only the bidirectionally consistent prototypes are kept and added to the pairing list. middle:

[0112]

[0113] After obtaining the prototype pairing list Then, pseudo-labels can be assigned to known category samples in the target domain according to the matching relationship. For unmatched target prototypes, the category is labeled as needed after the largest known category label number, and the corresponding samples are labeled with the same pseudo-label.

[0114] In this embodiment of the invention, the predicted classification label of the radiation source individual is output by the nearest neighbor classifier. The prediction method is to calculate the geodesic distance between the low-dimensional manifold feature input to the classifier and all low-dimensional manifold features in the classifier reference set, and select the category corresponding to the low-dimensional manifold feature in the reference set that is closest to it as the predicted output of the classifier.

[0115] In step 4, the source domain labeled low-dimensional features and the target domain pseudo-labeled low-dimensional features are used together as the reference set for the nearest neighbor classifier. The geodesic distance on the Grassmann manifold is used as the distance metric for the nearest neighbor classifier. Using this nearest neighbor classifier, the identification of radiation source individuals of known categories and different unknown categories in the target domain can be achieved.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0117] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources in open environments based on manifold learning, characterized in that, Includes the following steps: Step 1: Collect radiation signal data of various known radiation sources over a period of time using signal receiving equipment, and preprocess the collected radiation signal data to obtain multiple data samples of each radiation source and the radiation source category label of each data sample, so as to obtain the source domain dataset. In an open environment, radiation signal data of multiple radiation sources over a period of time are collected using signal receiving equipment. The collected radiation signal data are then preprocessed to obtain multiple data samples for each radiation source, thus obtaining a target domain dataset. Step 2: Construct an RF fingerprint feature extraction network to extract deep RF fingerprint features from the radiated signal, and train the feature extraction network using the source domain dataset; Step 3: Use the radio frequency fingerprint feature extraction network trained in Step 2 to extract deep radio frequency fingerprint features from the source domain dataset and the target domain dataset respectively, and reduce the dimensionality of the deep radio frequency fingerprint features to a low-dimensional Grassmann manifold to obtain the low-dimensional manifold features of the source domain and the target domain. The low-dimensional manifold features of the source domain are then mapped to the manifold of the target domain. Based on the geodesic distance on the Grassmann manifold, the low-dimensional manifold feature prototypes of the source and target domains are matched, and pseudo-labels are added to the target domain data according to the matching results. Step 4: Use the source domain labeled data and the target domain pseudo-labeled data together as the reference set for the nearest neighbor classifier, and predict the individual identification of radiation sources in the open environment through the nearest neighbor classification strategy.

2. The method as described in claim 1, characterized in that, Data preprocessing includes: extracting M seconds of radiation source signal data as a data sample, and normalizing the extracted signal data to linearly map the data to the [0,1] interval.

3. The method as described in claim 1, characterized in that, In step 2, during the training of the radio frequency fingerprint feature extraction network, a classifier is connected to predict the individual category label of the radiation source, and the radio frequency fingerprint feature extraction network is trained based on this classifier.

4. The method as described in claim 3, characterized in that, The classifier used to train the radio frequency fingerprint feature extraction network employs a single-layer fully connected layer structure.

5. The method as described in claim 1, characterized in that, The radio frequency fingerprint feature extraction network consists of a cascaded convolutional network module, several residual network modules, and an average pooling layer.

6. The method as described in claim 1, characterized in that, Step 3, which involves reducing the dimensionality of deep radio frequency fingerprint features to a low-dimensional Grassmann manifold, specifically includes: Calculate the autocorrelation matrix of the deep radio frequency fingerprint features in the source / target domain, and then perform singular value decomposition on it; Take the first left singular matrix obtained from singular value decomposition The column yields a low-dimensional feature representation of the data samples on the Grassmann manifold; where, This is a preset value.

7. The method as described in claim 1, characterized in that, Step 3 also includes: Set the optimal transport model for the source and destination domains on the Grassmann manifold: ； in, and These represent the source and target domain data, respectively. , These represent the feature indices of the source and target domains, respectively. This represents the optimal transmission plan from the low-dimensional feature distribution of the source domain to the low-dimensional feature distribution of the target domain. This represents a transport pair between the source and destination domains. Let f(x) represent the mapping function from low-dimensional features of the source domain to the manifold of the target domain. The low-dimensional features of the source domain after mapping These are the corresponding low-dimensional features of the source domain after optimal mapping. For transmission costs, Represents a transmission plan from a low-dimensional feature distribution in the source domain to a low-dimensional feature distribution in the target domain. entropy, These are preset weight parameters used to control the transmission plan. sparsity; Based on the solution results of the optimal transmission model, the low-dimensional manifold features of the source domain are mapped to the target domain manifold, and the mapped low-dimensional features of the source domain are obtained.

8. The method as described in claim 7, characterized in that, Transmission cost Specifically set as follows: ； in, This represents the low-dimensional features of the target domain of data sample j. Represents the low-dimensional features of the source domain after mapping of data sample i; The transmission cost from the low-dimensional feature distribution of the source domain to the low-dimensional feature distribution of the target domain. This can be understood as all known categories. The maximum value; For unknown categories, the penalty coefficient is... is the geodesic distance on the Grassmann manifold.

9. The method as described in claim 8, characterized in that, Unknown category penalty coefficient Among them, the minimum reconstruction error , The low-dimensional feature prototype of the source domain after mapping. This represents the low-dimensional feature prototype of the source domain after mapping of data sample i. For source domain category Category radius, source domain category low-dimensional features of the target domain The category of the nearest mapped source domain prototype, where any category Category radius Defined as: , The average geodesic distance from the low-dimensional feature within the class to the prototype of the low-dimensional feature. The standard deviation of the distance. A coefficient used to control the coverage area of ​​the radius.

10. The method as described in claim 1, characterized in that, In step 3, the low-dimensional manifold feature prototypes of the source and target domains are matched based on the geodesic distance on the Grassmann manifold, and pseudo-labels are assigned to the target domain data according to the matching results. Specifically, this includes: Calculation of low-dimensional manifold features for different categories of samples in the source domain dataset The mean is used to obtain the prototype of the low-dimensional manifold features in the source domain. ; The low-dimensional manifold feature prototype of the target domain is obtained by spectral clustering on the Grassmann manifold. ; Cross-domain prototype matching is performed based on the nearest neighbor strategy. The matching process is as follows: For each source domain prototype Find the target domain prototype with the closest geodesic distance in the target domain. ,in, This represents the index of the data sample found in the target domain; For each target domain prototype Similarly, find the source domain prototype with the closest geodesic distance in the target domain. ,in, This represents the index of the data sample found in the source domain; Keep the bidirectional consistent prototypes in the pairing list middle: in, This indicates a bidirectional, consistent prototype pairing; Based on the pairing list The matching relationship in the target domain is used to label pseudo-labels for data samples of known categories; for unmatched target domain prototypes, the category is labeled as needed after the largest known category label number, and pseudo-labels are labeled for the corresponding data samples based on the currently labeled category.

Images