A multi-platform multi-label radiation source individual identification method, system, device, medium and product

Abstract

The application discloses a multi-platform multi-label radiation source individual identification method, system, device, medium and product, and relates to the technical field of target identification. The method comprises the following steps: constructing a label based on radar platforms and target number data; obtaining a joint embedding vector based on the label and constructing a symmetric normalized adjacency matrix; obtaining sample radar data; obtaining sample deep residual signal features based on the sample radar data by using a deep residual network; obtaining efficient features based on the sample deep residual signal features and the joint embedding vector; obtaining a classifier based on the efficient features and the symmetric normalized adjacency matrix by using a GCN network; obtaining radiation source signal data in a set region and obtaining deep residual signal features by using a deep residual network; obtaining a score vector based on the deep residual signal features and the classifier; and obtaining radar platform numbers and target numbers of the radiation source signal data based on the score vector. The application can improve the identification speed and accuracy of the radiation source signal data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target identification, in particular to a multi-platform multi-label radiation source individual identification method, system, device, medium and product. BACKGROUND

[0002] The individual characteristics of radar radiation sources are formed by parasitic modulation generated by key devices of the transmitter (such as the transmitting tube, power amplifier, etc.). These characteristics cannot be avoided, eliminated and forged, and are unique due to differences in modulation form and modulation amount, thereby making individual identification possible. Specific Emitter Identification (SEI) is a technology that realizes the unique identification of radio radiation sources by measuring the external characteristics of intercepted electromagnetic signals. This technology is widely used in radar target detection, perception and other fields.

[0003] Existing SEI methods mainly fall into two categories: methods based on hand-crafted features and methods based on deep learning. Methods based on hand-crafted features extract features such as time domain, frequency domain, and time-frequency domain of signals, and use classifiers such as Support Vector Machine (SVM) and decision trees to complete the classification task. However, this type of method has a high dependence on expert knowledge and poor adaptability to complex electromagnetic environments. In contrast, methods based on deep learning automatically learn signal features to identify radiation source individuals, and their performance is superior to methods based on hand-crafted features. Common deep learning models include Convolutional Neural Network (CNN), Residual Network (ResNet), complex neural network, Gated Recurrent Unit (GRU), and Transformers model.

[0004] However, the above two types of SEI methods mainly classify and identify multiple targets based on a single carrier platform of a radar. Single-platform classification and identification may not be able to fully cover the target area due to limited data sources, and may be affected by local environmental factors (such as terrain obstruction and electromagnetic interference), resulting in decreased classification and identification accuracy, poor robustness, and a range of perception that cannot meet actual needs. Therefore, multi-platform distributed fusion identification technology has emerged. Currently, multi-platform distributed radiation source individual identification faces differences in heterogeneous feature distribution such as land-based complex electromagnetic interference, signal time variation caused by sea-based dynamic sea conditions, and air-based Doppler shift and limited airborne processing. Traditional homogeneous fusion methods cannot effectively train the model to converge and limit the identification performance due to the mismatch of multi-platform features. Therefore, how to use multi-platform distributed fusion to identify radiation source individuals in a feature-heterogeneous scene is a problem that needs to be solved.

[0005] There are currently several ways to solve the above problems: one is a multi-label classification method; two is a receiver dyeing effect compensation method; three is a transmitter fingerprint mapping modeling method. Since removing the receiver dyeing effect requires complete multi-channel full pulse data for training and debugging, the data collected in the actual measurement environment is limited, which limits the application of this method. In addition, due to the nonlinear time-varying characteristics of radar signals in complex and diverse application scenarios, it is difficult to accurately represent the differences between individual emitters by constructing a mapping function of the transmitter fingerprint under multi-scene migration conditions, so this method has certain limitations. Compared with the above, the multi-label classification method realizes feature decoupling by constructing a multi-dimensional label system, and does not depend on the integrity of multi-channel data and the consistency of signal distribution, which has a significant advantage in SEI problems with large differences in heterogeneous feature distribution.

[0006] Multi-label feature recognition is to use deep learning technology to effectively extract features from signal features and perform multi-label classification to achieve end-to-end recognition. In recent years, some existing researches have made a series of progress in multi-label feature recognition around three dimensions of feature representation optimization, model lightweight and cross-domain adaptability, but these existing researches still have fundamental limitations: the single-platform multi-label classification method models the "platform-target" combination as independent discrete labels, which leads to an exponential growth of the label space with the number of platforms, and requires a separate design of a feature extractor for each label. This not only ignores the associated dependency relationship between platforms and targets, but also introduces label noise interference due to unrelated and weakly related label combinations, resulting in waste of computing resources and decline in classification performance, and cannot adapt to the label space expansion needs of multi-platform system identification. SUMMARY

[0007] The purpose of the present application is to provide a multi-platform multi-label emitter individual identification method, system, device, medium and product, which can improve the identification speed and accuracy of emitter signal data.

[0008] To achieve the above purpose, the present application provides the following solutions:

[0009] In a first aspect, the present application provides a multi-platform multi-label emitter individual identification method, comprising:

[0010] Obtaining the number data of all radar platforms and the number data of all targets in a set area;

[0011] Constructing a label based on the number data of the radar platforms and the number data of the targets; obtaining a joint embedding vector based on the label; constructing a symmetric normalized adjacency matrix based on the label;

[0012] acquire radar data of all the targets received by all the radar platforms in a set region as sample radar data;

[0013] obtain sample deep residual signal features based on the sample radar data by using a deep residual network;

[0014] obtain efficient features based on the sample deep residual signal features and the joint embedding vector;

[0015] obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using a GCN network, and use the weight matrix as a classifier;

[0016] acquire radiation source signal data in a set region; the radiation source is a target in the set region;

[0017] obtain deep residual signal features based on the radiation source signal data by using the deep residual network, obtain a score vector based on the deep residual signal features and the classifier, and obtain radar platform numbers and target numbers of the radiation source signal data based on the score vector.

[0018] In a second aspect, the present application provides a multi-platform multi-label radiation source individual identification system, comprising:

[0019] a data acquisition module, configured to acquire number data of all radar platforms and number data of all targets, sample radar data, and radiation source signal data in a set region;

[0020] a label dependency modeling module, configured to construct labels based on the number data of the radar platforms and the number data of the targets, obtain a joint embedding vector based on the labels, and construct a symmetric normalized adjacency matrix based on the labels;

[0021] a feature representation learning module, configured to obtain sample deep residual signal features based on the sample radar data by using a deep residual network, obtain efficient features based on the sample deep residual signal features and the joint embedding vector, and obtain deep residual signal features based on the radiation source signal data by using the deep residual network;

[0022] a GCN classifier learning module, configured to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using a GCN network, and use the weight matrix as a classifier;

[0023] a data output module, configured to obtain a score vector based on the deep residual signal features and the classifier, and obtain radar platform numbers and target numbers of the radiation source signal data based on the score vector.

[0024] In a third aspect, the present application provides a computer device, comprising: a memory, a processor to store a computer program on the memory and run the computer program on the processor, and the processor executes the computer program to implement the steps of the multi-platform multi-label radiation source individual identification method according to any one of the above.

[0025] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the multi-platform multi-label radiation source individual identification method according to any one of the above.

[0026] In a fifth aspect, the present application provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the steps of the multi-platform multi-label radiation source individual identification method according to any one of the above.

[0027] According to the specific embodiments provided by the present application, the present application has the following technical effects:

[0028] The present application provides a multi-platform multi-label radiation source individual identification method, system, device, medium and product, constructs a label based on the number data of all radar platforms and the number data of all targets in a set area, constructs a joint embedding vector and a symmetric normalized adjacency matrix between the radar platforms and the targets according to the label, only retains the effective targets that can be detected by each radar platform, eliminates the influence of irrelevant label combinations on radiation source individual identification, and can improve the correlation dependence relationship between the radar platforms and the targets, improve the utilization rate of computing resources and identification performance. The deep residual network is used to obtain sample deep residual signal features based on sample radar data, to realize extraction of deep signal features, and to obtain efficient features based on the deep residual signal features and the joint embedding vector, and the GCN network is used to process the efficient features and the symmetric normalized adjacency matrix to obtain a classifier. The classifier is obtained by training the GCN network through a combination loss function, and is obtained based on the efficient features and the symmetric normalized adjacency matrix using the trained GCN network, which can solve the long-tail distribution problem of labels, improve the identification robustness of small sample categories, and thus meet the label space expansion demand of systematic identification covering multiple radar platforms. In actual application, the deep residual network is used to obtain deep residual signal features based on the obtained radiation source signal data in a set area, and to obtain a score vector based on the deep residual signal features and the classifier, and then to obtain the radar platform number and the target number of the radiation source signal data, thereby improving the identification speed and accuracy of the radiation source signal data. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0030] Figure 1 A flowchart of a multi-platform multi-label radiation source individual identification method in an embodiment of the present application;

[0031] Figure 2 A framework diagram of the multi-platform multi-label radiation source individual identification method provided by an embodiment of the present application;

[0032] Figure 3 A whole flowchart of the multi-platform multi-label radiation source individual identification method provided by an embodiment of the present application;

[0033] Figure 4 A whole architecture diagram of the feature representation part provided by an embodiment of the present application;

[0034] Figure 5 A structure diagram of the multi-platform multi-label radiation source individual identification system provided by an embodiment of the present application;

[0035] Figure 6 A schematic diagram of the relative layout positions of the radar platform and the target provided by another embodiment of the present application;

[0036] Figure 7 A schematic diagram of the parameter setting and tuning results of the combined loss function provided by another embodiment of the present application;

[0037] Figure 8 A comparison diagram of the influence of different GCN layers on the identification accuracy provided by another embodiment of the present application;

[0038] Figure 9 A comparison diagram of the influence of different dimension transformation strategies on the identification accuracy provided by another embodiment of the present application;

[0039] Figure 10 A comparison diagram of the identification accuracy under three kinds of loss functions provided by another embodiment of the present application;

[0040] Figure 11 A confusion matrix diagram of the multi-platform multi-label radiation source individual identification method and the traditional CNN multi-classification network model provided by another embodiment of the present application;

[0041] Figure 12 A performance comparison result diagram provided by another embodiment of the present application;

[0042] Figure 13 A biased test confusion matrix diagram is provided for another embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0044] The above purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be further described in detail below with reference to the drawings and specific embodiments.

[0045] In an exemplary embodiment, as shown in Figure 1 , Figure 2 and Figure 3 , a multi-platform multi-label radiation source individual identification method is provided, comprising:

[0046] Step 100, obtain the number data of all radar platforms and the number data of all targets in a set area. Construct a label based on the number data of the radar platforms and the number data of the targets.

[0047] Step 200, obtain a joint embedding vector based on the label. Construct a symmetric normalized adjacency matrix based on the label.

[0048] Step 300, obtain all target radar data received by all radar platforms in the set area as sample radar data.

[0049] Step 400, use a deep residual network to obtain a sample deep residual signal feature based on the sample radar data. Obtain an efficient feature based on the sample deep residual signal feature and the joint embedding vector.

[0050] Step 500, use a GCN network to obtain a weight matrix based on the efficient feature and the symmetric normalized adjacency matrix, and use the weight matrix as a classifier.

[0051] Step 600, obtain radiation source signal data in the set area. The radiation source is a target in the set area. Use a deep residual network to obtain a deep residual signal feature based on the radiation source signal data. Obtain a score vector based on the deep residual signal feature and the classifier. Obtain the radar platform number and the target number of the radiation source signal data based on the score vector.

[0052] Wherein, the radar platform refers to a radar signal receiver platform, which can receive radar data (i.e. radar signal data) of a target (i.e. radiation source).

[0053] As an optional implementation, in order to improve the performance of individual identification of the radiation source, the process of obtaining the joint embedding vector based on the label in step 200 includes: obtaining a one-hot vector based on the label. The embedding vector of the radar platform and the embedding vector of the target are obtained based on the one-hot vector. The joint embedding vector is obtained based on the embedding vector of the radar platform and the embedding vector of the target.

[0054] For example, in actual radar detection applications, radar signal data usually comes from multiple radar platforms, containing radar data of different targets. Implicit modeling can enhance the expression ability of the label, and as a label feature and data feature are spliced, it is convenient for the classifier to fuse the label feature and the signal feature to improve the classification and identification performance.

[0055] In the implicit modeling of label dependence, considering radar signal data from any radar signal receiver platform and for any target, the label of the “platform-target” combination is defined and constructed according to the obtained radar platform number data and target number data, denoted as , wherein and respectively represent the number of radar platforms and the number of targets, which are used to represent the source platform and the corresponding target of the radar data. The one-hot vector obtained based on the label is shown in formula (1).

[0056] (1)

[0057] In the formula, represents the one-hot vector, and respectively represent the one-hot encoding of the radar platform and the one-hot encoding of the target, . and respectively represent real number vector spaces with dimensions and . and are the total number of radar platforms and the total number of targets, respectively.

[0058] In order to improve data operation efficiency while preserving key information for easy extraction, the high-dimensional one-hot vector is mapped to a low-dimensional space through a shared embedding matrix to obtain the embedding vector of the radar platform and the embedding vector of the target, and then the joint embedding vector is obtained based on the embedding vector of the radar platform and the embedding vector of the target, as shown in formulas (2) and (3).

[0059] (2)

[0060] (3)

[0061] In the formula, These represent the shared embedding matrix of the radar platform and the shared embedding matrix of the target, respectively, to preserve the semantic structure of all labels. It is the embedded dimension. and They represent dimensions as follows: and The space of real matrix numbers. These represent the embedding vector of the radar platform, the embedding vector of the target, and the joint embedding vector, respectively.

[0062] In explicit modeling of label dependencies, the complex dependency structure between labels is captured by constructing associations between them. A correlation matrix based on co-occurrence patterns is designed to guide information propagation between nodes in the GCN network. To ensure numerical stability and explicitly model sub-links, a co-occurrence matrix is ​​defined. As shown in formula (4).

[0063] (4)

[0064] in, Co-occurrence matrix The element in the text is used to quantize the label. and tags The strength of the association and the co-occurrence frequency between them Indicates label and The statistical frequency (or similarity) of co-occurrence in the data. This is an optional, tiny positive number used to prevent the diagonal elements of the co-occurrence matrix from being zero. This refers to the Kronecker delta function, which ensures that each label... and The self-join weight is at least [value missing]. This avoids numerical instability or graph structure breakage caused by zero co-occurrence frequency. Satisfying formula (5):

[0065] (5)

[0066] To construct label dependencies by explicitly assigning self-joins and neighbor weights, a weighted adjacency matrix is ​​built. As shown in formula (6).

[0067] (6)

[0068] in, Represents the weighted adjacency matrix The Middle Line number The elements of the column are used to quantize the labels. and the connection weight between label . is the preset self-connection weight, usually set to 0.8, representing the dependence of label node on its own features. is an optional tiny regularization term to ensure numerical stability. The weighted adjacency matrix The weights of the non-diagonal elements in the weighted adjacency matrix are proportionally assigned according to the co-occurrence frequency, with a total neighbor weight of , usually set to 0.2, implying the strength of the inter-label dependency, while balancing the problems of overfitting and oversmoothing. represents the element in the th row and the th column of the co-occurrence matrix , used to quantify the association strength between label and label .

[0069] To construct the symmetric normalized adjacency matrix , for any node , formula (7) should be satisfied:

[0070] (7)

[0071] where represents the element in the th row and the th column of the symmetric normalized weighted adjacency matrix , used to quantify the normalized connection weight between label and label . represents the diagonal element of the corresponding degree matrix of the symmetric normalized adjacency matrix, representing the normalized degree of label , used to depict the sum of connection strength of label in the graph structure, supporting the symmetric normalization operation. represents the element in the th row and the th column of the augmented matrix , obtained by adding the elements in the corresponding positions of and the identity matrix , used to explicitly retain the self-connection weight, ensuring the integrity of the graph structure, and serving as an intermediate transition for the symmetric normalization operation. represents the element in the th row and the th column of the identity matrix with dimension , used to construct the augmented matrix , the unit matrix diagonal elements are 1, and the rest are 0, which is used to force each label to add self-connection, avoid abnormal graph convolution calculation due to no self-connection of the label, and strengthen the preservation of the label's own characteristics.

[0072] In terms of label dependency modeling, the implicit modeling module of multi-level label embedding establishes a unified semantic feature representation of heterogeneous labels, maps the discrete platform and target numbers to continuous label vectors with certain relevance, and defines them as label features. Then, through explicit modeling, a sparse correlation matrix is constructed and a matrix pruning strategy is implemented to construct a symmetric normalized adjacency matrix, effectively eliminating redundant label interference. At the same time, a reweighting mechanism is introduced to adjust the weight of the graph structure, which not only maintains the accurate expression of label relevance, but also suppresses the over-smoothing phenomenon in the graph convolution process, laying a foundation for subsequent cross-modal feature interaction.

[0073] As an optional implementation, in order to facilitate subsequent feature extraction, the sample radar data in the set region obtained in step 300 and the radiation source signal data obtained in step 600 are all preprocessed data. The preprocessing process specifically includes: performing Fourier transform on the radar signal data to obtain complex frequency domain components of the radar signal data. The mapping features are obtained based on the amplitude spectrum of the complex frequency domain components. The mapping features are linearly scaled to obtain the preprocessed radar signal data (such as sample radar data, radiation source signal data).

[0074] For example, for the scenario where a fixed receiver (i.e., a radar platform) works with a single radiation source (i.e., a target), the signal modeling process can be achieved through nonlinear device characteristic analysis. Under this timing separation receiver mechanism, the separation requirement of multiple radiation source signals is avoided, as the physical layer characteristics mainly come from the nonlinear distortion effect of the power amplifier (PA). Let the radio frequency modulated signal input to the power amplifier be represented by formula (8):

[0075] (8)

[0076] wherein, represents the time domain sampling value of the radio frequency signal after the baseband modulated signal is modulated by the carrier wave, and in represents the input after modulation. represents the baseband modulation waveform, and correspond to the carrier frequency and sampling rate parameters, respectively. is the time sampling index, which is used to convert continuous signals into discrete signals for processing.

[0077] In order to characterize the nonlinear transmission characteristics of the power amplifier, a polynomial series expansion model is used, and the output signal of the power amplifier in a single-hop communication scenario can be represented by formula (9):

[0078] (9)

[0079] wherein, denotes the power amplifier output signal of the thradiation source, is the polynomial expansion coefficient (i.e., series expansion coefficient) of the power amplifier of the thradiation source, different coefficients represent the power amplifier difference between different radiation source individuals. K represents the non-linear order, denotes the number of radiation sources.

[0080] The discrete sampling signal of the thradiation source acquired by the receiver (i.e., radar platform) satisfies formula (10):

[0081] (10)

[0082] wherein, denotes the channel fading factor from the transmitting end of the thradiation source to the receiver, is the channel additive noise term of the thradiation source, is the sampling block index of the received signal.

[0083] Further expanding formula (10) can obtain the analytical expression of the received signal, as shown in formula (11).

[0084] (11)

[0085] The above steps quantize the power amplifier distortion effect through nonlinear series expansion, and simultaneously model the channel attenuation and noise interference. The heterogeneous radiation source signal is converted into a feature vector with unified dimension and comparability, which facilitates the realization of cross-sample consistent representation. In order to facilitate subsequent feature extraction, the signal needs to be further transformed in the signal domain and statistically standardized.

[0086] The time-domain signal is subjected to Fourier transform at W points to obtain the frequency domain representation, as shown in formula (12).

[0087] (12)

[0088] wherein, is the complex frequency domain component of the thradiation source, and its amplitude spectrum is usually taken as the analysis object. Wherein, is the frequency domain sampling index, representing the thfrequency point. is the imaginary unit.

[0089] To suppress the difference in the dynamic range of the spectrum, a logarithmic transformation and offset compensation are applied to the amplitude spectrum of the complex frequency domain components, as shown in formula (13).

[0090] (13)

[0091] in, This is used to avoid numerical singularities caused by zero-value inputs, while also enhancing the distinguishability of low-energy frequency bands. Indicates the first The amplitude spectrum characteristics of the complex frequency domain components of a radiation source after logarithmic transformation and offset compensation. This is the frequency domain sampling index.

[0092] Based on the statistical properties of all samples (i.e., the sampled signals), the features are mapped to the standardized interval, as shown in formula (14).

[0093] (14)

[0094] in, Indicates the first Amplitude spectrum characteristics after normalization of individual radiation sources and Let represent the mean and variance of all samples, respectively. is the numerical stability constant.

[0095] Ultimately, through linear scaling, Constrain to the range (0,1), as shown in formula (15).

[0096] (15)

[0097] in, Indicates the first The amplitude spectrum characteristics of a radiation source after linear scaling. These represent the maximum and minimum values ​​of the amplitude spectrum characteristics of all radiation sources, respectively.

[0098] As an optional implementation, step 400, which obtains efficient features based on sample depth residual signal features and joint embedding vector, includes: using attention weighting to concatenate the sample depth residual signal features and joint embedding vector to obtain efficient features.

[0099] For example, the sample radar data is first dimensioned by signal reconstruction, a deep residual network architecture is adopted to stack a jump connection module, and a multi-level residual learning is used to synchronously capture local detail features and global feature information of the signal. Based on the sample deep residual signal features extracted by the joint embedding vector and the deep residual network, the two are fused in a splicing layer, and then the feature dimension selection is implemented through a double-channel attention weighting module, specific weights are applied to the feature channels of the signal domain and the label domain respectively, and efficient feature representation is realized. The overall architecture of the feature representation part is as shown in Figure 4

[0100] ResNet is used as the basic network model, the deep structure of ResNet-101 network is adopted, the signal features of the sample radar data after dimension transformation are extracted through convolution layer and residual block, and the sample deep residual signal features are obtained by performing average pooling and dimension transformation processing on the extracted signal features. Further, the attention weighting is used to splice the sample deep residual signal features and the joint embedding vector to obtain efficient features. As shown in equation (16).

[0101] (16)

[0102] In the formula, represents the sample radar data, represents the efficient feature, represents the sample deep residual signal feature, represents the attention weighting of the feature after splicing and fusion of the sample deep residual signal feature and the joint embedding vector feature, represents the dimension conversion processing, represents the feature extraction using CNN convolutional neural network, represents the average pooling operation, represents the attention weighting mechanism, represents the dimension of the efficient feature, represents represents the efficient feature vector space.

[0103] As an optional implementation, in order to improve the accuracy of identification, step 500 includes: using a GCN network to obtain an initial classifier based on the efficient feature and the symmetric normalized adjacency matrix. An initial score vector is obtained based on the initial classifier and the sample deep residual signal feature, and an initial identification result is obtained based on the initial score vector. The GCN network is trained based on the initial identification result and the true result using a combined loss function until the loss function value meets the set requirement, and the classifier is obtained based on the efficient feature and the symmetric normalized adjacency matrix using the trained GCN network.

[0104] In the formula, the combined loss function is shown in equation (17).

[0105] ​ (17)

[0106] In the formula, This represents the loss value of the initial recognition result. Indicates the number of radar platforms. Indicates the number of targets. Indicates the first in the true result The actual value of each element Indicates the initial recognition result for the first The probability that an element is predicted to be 1. This indicates the composite weight.

[0107] For example, due to differences in detection and stealth capabilities among various radar platforms and targets, there are certain combinational relationships between radar platforms and the detected targets. Simply simplifying the multi-label classification problem mathematically to a binary classification problem often ignores the correlation structure between radar platforms and targets, causing the problem analysis to remain at the level of formulaic structural reasoning. When the number of platform and target categories is large, there will be difficulties in high computational complexity and scalability, thus being fundamentally limited. The idea of ​​recognizing and classifying multiple targets in samples based on probabilistic graphical models in GCN networks is applied to the one-dimensional radar signal feature extraction problem. It explicitly models the label correlation between radar platform categories and target categories, which can greatly save computational resources, achieve rapid analysis, and effectively avoid the impact of unrelated or weakly related pairs between "platform-target" on the accuracy of predicted categories. In GCN networks, node features are represented by adjacency matrices. To disseminate information, each node The feature update is shown in formula (18).

[0108] (18)

[0109] in, Indicates label After the first Feature representation after processing by the layered GCN network For the first The weight matrix of the learnable parameters of the layer GCN network. For activation functions, neighbor terms Neighbor information is aggregated by weighting co-occurrence frequency.

[0110] go through The result obtained after processing by the layer GCN network is shown in formula (19).

[0111] (19)

[0112] in, The feature representation (i.e., efficient features) initially input into the first layer of the GCN network. ), For the process Output results of the layer GCN network Indicates the first The bias term in the layered GCN network is used to adjust the network output and increase the expressive power of the network model.

[0113] The initial score vector is obtained based on the initial classifier (i.e., the output of the last layer of the GCN network) and the sample depth residual signal features. The initial recognition result is obtained based on the initial score vector. The GCN network is trained based on the initial recognition result and the real result using the combined loss function shown in formula (17) until the loss function value reaches the set requirement. The classifier is obtained based on the efficient features and the symmetric normalized adjacency matrix using the trained GCN network.

[0114] For example, in multi-label classification tasks, especially in the problem of identifying individuals from multiple platforms and multiple labels, class imbalance and modeling dependencies between labels are two key issues. Traditional loss functions (such as cross-entropy loss) perform poorly when handling imbalanced data and cannot explicitly model dependencies between labels. Therefore, a combined loss function (CB) is designed, which integrates focus loss (FL), multi-label loss (ML), and imbalanced category loss (IC). By dynamically adjusting the loss weights, the contribution of minority class samples is increased; by explicitly modeling the correlation between labels, the complex dependency structure between labels is captured; and through joint optimization, the classification performance of the model on multi-label data is improved.

[0115] Since multi-label classification is involved, the multi-label loss function MLL is chosen as the base function, as shown in formula (20).

[0116] (20)

[0117] in, This represents the value of the multi-label loss function. , .

[0118] To achieve the goal of dynamically adjusting and balancing the weights of different classes, suppressing the gradient contribution of easy samples (high confidence), and focusing on learning from difficult samples (low confidence), FL weights are introduced. Inverse Frequency (IF) Weights and dynamic index adjustment items As shown in formula (21).

[0119] (twenty one)

[0120] Among them, when When it approaches 1, Approaching 0 reduces the loss contribution of simple samples; when When approaching 0, Increase the size and strengthen the learning of difficult samples. The first in the actual results The number of times each element is not zero. The control function focuses more on difficult-to-classify samples; and To control the shape of the weight decay curve and suppress the weights of high-confidence samples; It is a smoothing term.

[0121] Combining the three factors in formula (21), we obtain the composite weight. As shown in formula (22).

[0122] (twenty two)

[0123] Composite weight The loss term applies only to positive samples, while negative samples retain their original weights, forming the final combined loss function. Formula (17) is expanded in combination with formula (22) to obtain formula (23).

[0124] (twenty three)

[0125] As an optional implementation, in order to improve classification performance, the process of obtaining the score vector based on the deep residual signal features and the classifier in step 600 includes: performing convolutional fusion on the deep residual signal features and the classifier to obtain the score vector.

[0126] For example, the features of the deep residual signal are convolved and fused with the graph structure features (i.e., the classifier, which is a weight matrix containing several parameters) of the final output of the trained GCN network to obtain a score vector. As shown in formula (24).

[0127] (twenty four)

[0128] in, This represents the result of the last layer of GCN operation (i.e., the classifier). This represents the characteristics of the deep residual signal.

[0129] Furthermore, to ensure that the combined loss function supports end-to-end optimization, it is necessary to verify the differentiability of the combined loss function with respect to the parameters. Therefore, for We verified that the gradients of both the positive and negative sample parts of the combined loss function are differentiable.

[0130] For positive samples ( ) and negative samples ( As shown in formulas (25) and (26), the derivative of the loss function with respect to the positive and negative sample terms yields:

[0131] (25)

[0132] (26)

[0133] The total gradient of the combined loss function is the weighted sum of the gradients of the positive and negative samples, as shown in formula (27).

[0134] (27)

[0135] As can be seen from the above equation, the hyperparameters in the combined loss function are logarithmic functions. and exist Internally differentiable; exponential function Infinitely differentiable in any real number field; polynomials and power functions and exist It is intrinsically differentiable; the product of all derivative terms and the linear combination thereof are differentiable, therefore the loss function... right The gradient exists everywhere and is continuous, satisfying the differentiability requirement for backpropagation.

[0136] Based on the same inventive concept, this application also provides a multi-platform, multi-label radiation source individual identification system. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations in the following embodiments of the multi-platform, multi-label radiation source individual identification system can be found in the limitations of the multi-platform, multi-label radiation source individual identification method described above, and will not be repeated here.

[0137] In one exemplary embodiment, such as Figure 5 As shown, a multi-platform, multi-label radiation source individual identification system is provided, including:

[0138] The data acquisition module is used to acquire the identification data of all radar platforms and all targets within a set area, sample radar data, and radiation source signal data within the set area.

[0139] The label dependency modeling module is used to construct labels based on the radar platform's identification data and the target's identification data, obtain a joint embedding vector based on the labels, and construct a symmetric normalized adjacency matrix based on the labels.

[0140] The feature representation learning module is used to obtain sample depth residual signal features based on sample radar data using a deep residual network. It is also used to obtain efficient features based on sample depth residual signal features and joint embedding vectors, and to obtain depth residual signal features based on radiation source signal data using a deep residual network.

[0141] The GCN classifier learning module is used to obtain the weight matrix based on efficient features and symmetric normalized adjacency matrix using the GCN network, and then uses the weight matrix as the classifier.

[0142] The data output module is used to obtain a fractional vector based on the deep residual signal features and the classifier, and to obtain the radar platform number and target number of the radiation source signal data based on the fractional vector.

[0143] In an exemplary embodiment, combining the multi-platform, multi-tag radiation source individual identification method described in the above embodiments, experimental analysis was conducted on the identification method using measured data from two radar signal receiver platforms (Platform 1 and Platform 2) for six targets (Target 1 to Target 6). The collected target radiation source signal data came from the same type of X-band maritime navigation radar, with LFM modulation. The radar parameters are shown in Table 1. The radar platform receiver used a portable ultra-wideband real-time spectrum analyzer with an instantaneous bandwidth of 80MHz.

[0144] Table 1 Radar Parameter Table

[0145]

[0146] To adapt to multi-platform, multi-target identification tasks, different combinations of platform targets were divided into 12 experimental groups (numbered 1 to 12). Platform 1 and target i (i=1-6) correspond to groups 1 to 6, and platform 2 and target i (i=1-6) correspond to groups 7 to 12. The radar signal receiver platform and six target radiation sources were deployed at designated locations (corresponding to platform 1 and platform 2, targets 1-6). The farthest and closest distances between the two platforms and the six targets were 21.8 km and 17.8 km, and 11 km and 6.3 km, respectively. The positions of the platforms and targets, as well as their surroundings, were not in waterways, had few ships, and were minimally affected by environmental interference. The relative deployment positions of the platforms and targets are as follows: Figure 6 As shown. The data acquisition scenario is set as follows: six target radars are powered on simultaneously, and the transmission power of each navigation radar is manually adjusted so that the signal of a single target radiation source is relatively strong. At this time, the receivers of two radar platforms identify the target radiation source with the strongest received signal.

[0147] Since the actual data collection did not involve combinations of some platforms and targets, the measured data only included groups 1-4, 7, and 10-12. To avoid the influence of time factors on the experimental results, all measured data without time offset were collected within one month. To further verify the robustness of the recognition method to time distribution shifts, a validation experiment with time offset was designed. For each existing group, data within six months of actual testing was used as the time offset test set to evaluate the model's applicability under time distribution conditions. Based on this, two classification tasks were constructed: a no-time-off scenario (data within one month) and a time-off scenario (data outside one month but within six months). The experimental design divided the dataset according to an 8:2 ratio, as shown in Table 2.

[0148] Table 2 Dataset partitioning table for two scenarios

[0149]

[0150] The above experiments were implemented using the PyTorch 2.0.1 framework and running on Windows 10 Professional. The hardware configuration included an Intel(R) Core™ i9-9900K CPU (3.6GHz) and an NVIDIA GeForce RTX A6000 GPU. Accuracy was used as the evaluation metric, with 20 training epochs. The initial learning rate was set to 0.00125, and cosine annealing was used to update the learning rate in each epoch. Regularization was introduced to prevent overfitting.

[0151] The design of the loss function is crucial for classifier optimization in recognition methods. In setting the parameters of the combined loss function, it is essential to determine the hyperparameters within the aforementioned combined loss function. , and To find the optimal value, this experiment uses a phased grid search method to fine-tune the parameters on an unbiased test set to balance prediction accuracy in class imbalance scenarios. First, based on the baseline values ​​determined in preliminary experiments, the focus loss adjustment factor is then fixed. and inverse frequency attenuation coefficient and momentum factor The two parameters in the model are used to search for five discrete values ​​of the third hyperparameter on the validation set, and the validation results are cross-validated to obtain the local optimum value of the third hyperparameter under the experimental conditions. The specific hyperparameter settings and results are as follows. Figure 7 As shown. Figure 7 Parts (a), (b), and (c) are respectively Time The result of seeking optimization Time The results of optimization and Time The result of the optimization. Based on... Figure 7 The results shown indicate that, under the current data and experimental conditions, when At that time, the model's accuracy on the unbiased test set improved to 95% in 5 epochs of training and gradually approached convergence, proving that the model achieved optimal performance when the parameter was set, and the optimization process was stable.

[0152] Based on message passing mechanism theory, the number of layers in a GCN network determines the range of neighborhood information aggregation for node features, directly affecting the model's ability to capture graph structure information and thus its classification ability. However, excessive layer stacking can lead to oversmoothing, causing node features to become homogenized and blurring classification boundaries. To quantitatively analyze the impact of the number of GCN layers in GCN network structure selection, comparative architectures with 1 to 5 layers were constructed. The same initialization strategy was used to control variables, and model performance was evaluated by observing classification accuracy. The results are as follows: Figure 8 As shown.

[0153] The results show that when the GCN has 2 layers, the model achieves a classification accuracy of over 95% only at epoch 5, while the classification accuracy of other GCN layers does not reach 90% at this point. Under the same training epochs, it achieves high classification accuracy while also exhibiting fast convergence capability, ultimately reaching a peak accuracy of 97%. The experimental results verify the effectiveness of the model structure design. They also demonstrate that deep GCNs suffer from feature degradation, meaning that as the number of convolutional layers increases, some features are excessively smoothed. The 2-layer GCN architecture achieves an optimal balance between local neighborhood information aggregation and global structure preservation, providing a theoretical basis for subsequent model design.

[0154] Regarding the choice of dimensionality transformation, the dimensionality transformation of the input data (each sample's concatenated feature is 1×512-dimensional) is crucial to the model's feature extraction capability. For the feature adaptation problem of one-dimensional radar signals in a two-dimensional convolutional architecture, based on the local perceptual field theory of convolutional neural networks, five groups of data dimensionality transformations were designed: (1, 512), (2, 256), (4, 128), (8, 64), and (16, 32). The classification accuracy of different groups was compared to select a suitable dimensionality recombination scheme for the model. Furthermore, based on its convergence speed, appropriate training truncation can be performed to save computational resources. The results are as follows: Figure 9 As shown.

[0155] The results show that all groups can achieve a recognition accuracy of over 97% after 20 training sessions. However, the 16×32 group strategy can achieve a classification accuracy of over 95% in 5 epochs, which is faster than other groups. This is because this dimension has better geometric adaptability to the residual convolutional feature extraction network. Through parameter sharing efficiency optimization and computational graph fusion, while maintaining a certain feature representation ability, the overhead of computational resources is reduced, which can achieve fast convergence.

[0156] Through the above comparative experimental parameter setting optimization experiment, it can be seen that selecting... With 2 GCN layers and a dimensionality transformation strategy of (16, 32), the model performance can achieve local optima. At the same time, since the model has a faster convergence capability under this setting, the number of training rounds can be set to 10 according to the computing power and the requirements for real-time operation.

[0157] To address the class imbalance problem in multi-platform, multi-label recognition tasks, this experiment compares the designed combined loss function (CB) with the class imbalance function (IC) and focus function (FL). The recognition accuracy of the method is evaluated by training the GCN network on the dataset shown in Table 1 with the locally optimal parameter settings selected above. The results are as follows: Figure 10 As shown in the figure. Experimental results show that, under the propagation optimization effect of the designed combined loss function, the recognition accuracy is higher than the other two, and the training time to reach a certain accuracy is shorter. This proves the effectiveness of the application of implicit label modeling and weight adjustment strategies in the designed combined loss function in multi-platform and multi-label scenarios, and it is also effectively applicable to the class imbalance problem.

[0158] To verify the advantages of graph structure classification models in modeling complex relationships, this experiment compares the performance of a multi-platform, multi-label radiation source individual identification method with a traditional CNN multi-classification network model, evaluating their differences in accuracy and time complexity. Traditional methods, based on the CNN classification strategy, treat the "platform-target" combination as 12 independent labels, constructing a high-dimensional fully connected classifier after the ResNet-101 feature extraction network module, and using fully connected layers for final classification, while ignoring the correlation between the platform and the target. The confusion matrix and computation time per training epoch of the identification method provided in this application compared to the CNN-based multi-classification model are shown below. Figure 11 Parts (a) and (b) of the document and Table 3 are shown in Table 3. and These represent the computation time of the identification method in this application and the traditional multi-classification model, respectively.

[0159] Table 3. Comparison of computation time complexity for each training round of the two methods.

[0160]

[0161] Depend on Figure 11 As can be seen from part (a) of the document, the overall recognition accuracy of the recognition method in this application can reach over 97%, and the recognition accuracy of each imbalanced category is over 94%; Figure 11 As shown in part (b), the overall recognition accuracy of the traditional multi-classification model is only 80%, and the recognition accuracy of each category is only above 74%. In this problem, the recognition performance is poor, and there is a 3% probability of misclassification for the three unrelated groups of categories 5, 8, and 9. The experimental results show that the traditional multi-classification model ignores the constraint relationship between the platform and the target (e.g., platform 1 can only detect specific targets), resulting in the correlation not being explicitly modeled, which introduces a certain amount of label noise. At the same time, the independent classification assumption of the fully connected layer ignores the hierarchical dependency between the platform and the target, and cannot capture the cross-label collaborative features, resulting in some invalid label combinations participating in the calculation and being misclassified, which brings an unknown impact to the final prediction result.

[0162] As shown in Table 3, the graph structure model (i.e., the classifier) ​​based on the GCN network, using an 8-node (2 platforms + 6 targets) bipartite graph design and employing an adjacency matrix to encode association strength, takes an average of 154.63 seconds per training round. In contrast, the traditional connection design based on CNN with 12 nodes (2 platforms * 6 targets) takes an average of 170.28 seconds per training round. These results demonstrate that the classification in this application takes less training time and has higher computational efficiency under these conditions and dataset.

[0163] To verify the advantages of graph structure classification models (i.e., classifiers) in modeling complex relationships, this experiment uses the recognition method provided in this application (i.e., the graph convolutional deep residual network for classification (GCDR-Net) which integrates GCN, deep residual networks, and multi-platform multi-label word embeddings), corresponding to... Figure 12 A performance comparison was conducted between GCDR-Net and a simplified recognition method that removes implicit modeling and performs feature concatenation operations. The differences in accuracy across different training epochs and categories were evaluated, with all other parameters remaining the same. The comparison results are shown below. Figure 12 As shown.

[0164] Experimental results show that the recognition method provided in this application achieves implicit modeling by embedding the "platform-target" combined label into the classifier mapping. During training, by sharing parameters, all classifiers retain their semantic structure, and the gradients of all classifiers influence the classifier generation function. This results in an average accuracy of 95% and 97% after 5 and 20 epochs, respectively, with minimum accuracies of 90% and 92% for each category. In contrast, the simplified recognition method without implicit label modeling and feature concatenation achieves only 85% and 89% average accuracy after 5 and 20 epochs, respectively, with minimum accuracies of 81% and 85% for each category. Therefore, under the given dataset and parameter settings, the introduction of implicit label modeling and feature concatenation strategies in this problem improves the overall classification accuracy and minimum accuracy of each category by 8% and 7% after 20 training epochs, respectively, thus enhancing feature representation capabilities.

[0165] Individual radiation source identification methods need to possess temporal generalization capabilities to address feature distribution shifts caused by factors such as equipment aging and environmental disturbances. Therefore, a pre-trained model, trained using data from the past month and validated using an unbiased test set, will be tested on a cross-time period test set (the biased test set shown in Table 2) to evaluate the model's robustness under time-drift scenarios. The resulting biased test confusion matrix is ​​shown below. Figure 13 As shown.

[0166] The experimental results show that the recognition method provided in this application can still maintain an overall accuracy of 89% over a six-month time drift, proving that the proposed scheme has a certain degree of non-uniform stability in time drift scenarios. Some categories can still maintain 100% and 99% recognition accuracy. However, due to the differences in the amount of time drift and the amount of training data for each category, the prediction accuracy of some groups does not reach 80% or is even only 65%. This is because, during the time evolution, the feature hyperplane supported by the limited training samples of some categories cannot fully adapt to the time-varying nonlinearity of the channel, resulting in increased sensitivity of the decision boundary. Therefore, a newer strategy needs to be designed to improve this.

[0167] Based on the comparative experiments and results analysis above, this application proposes a method for identifying individual radiation sources across multiple platforms and labels. By explicitly constructing a platform-target association matrix and jointly optimizing implicit word embeddings, the model effectively captures the complex dependencies between multiple labels. Implicit modeling of label features and signal feature fusion enhances feature representation capabilities. Furthermore, a lightweight GCN network and residual network are used to achieve hierarchical extraction and fusion of signal features. Experiments show that the proposed method achieves a classification accuracy of 97% in complex real-world noise environments, a 17% improvement over traditional multi-classification models without graph structures, and a 9.3% reduction in single-round training time, an 8% improvement over the proposed model without implicit label modeling and feature concatenation strategies, demonstrating both efficiency and robustness. In a six-month time-shift scenario, the model maintains 89% recognition accuracy, verifying its adaptability to feature distribution shifts. However, the performance degradation of individual categories over long-term evolution indicates the inadequacy of the dynamic feature drift compensation mechanism. Future research will explore online incremental learning and adaptive graph structure optimization strategies to further improve the model's temporal generalization ability. This application provides theoretical support and technical pathways for the intelligent upgrading of multi-platform collaborative sensing systems.

[0168] Based on the descriptions in the above embodiments, this application proposes a multi-platform, multi-label radiation source individual identification method to address the problems of large heterogeneous feature distribution differences and poor identification performance in the identification of individual radiation sources from multiple radar platforms. First, a "platform-target" correlation matrix is ​​constructed based on the statistical characteristics of radar data. Then, label features are implicitly modeled through word embedding, and deep signal features are extracted using a deep residual convolution module. Simultaneously, the data features and label features are attention-weighted and concatenated. Finally, to fully utilize label correlation, a graph convolutional neural network is improved, using the concatenated features as input to generate a feature map and a classifier. The feature extraction results are then fused to generate the final classification result. Experimental data shows that this method achieves dynamic fusion and efficient classification of multi-platform, multi-label features. Compared to traditional independent multi-classification methods, the accuracy and computational efficiency are improved by 17% and 9.3%, respectively. The introduction of implicit label modeling feature concatenation further improves accuracy by 8%, and the identification performance fully meets the application requirements of practical scenarios.

[0169] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0170] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0171] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0172] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0173] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-platform multi-label radiation source individual identification method, characterized in that, The method comprises the following steps: obtaining the number data of all radar platforms and the number data of all targets in a set region; constructing labels based on the number data of the radar platforms and the number data of the targets; obtaining joint embedding vectors based on the labels; constructing a symmetric normalized adjacency matrix based on the labels; obtaining all radar data of all targets received by all radar platforms in the set region as sample radar data; obtaining sample deep residual signal features based on the sample radar data by using a deep residual network; obtaining efficient features based on the sample deep residual signal features and the joint embedding vectors; obtaining a weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using a GCN network, and taking the weight matrix as a classifier; obtaining radiation source signal data in the set region; the radiation source is a target in the set region; obtaining deep residual signal features based on the radiation source signal data by using the deep residual network; obtaining a score vector based on the deep residual signal features and the classifier; and obtaining the number of radar platforms and the number of targets corresponding to the radiation source signal data based on the score vector.

2. The method according to claim 1, wherein The method for obtaining joint embedding vectors based on the labels comprises the following steps: obtaining one-hot vectors based on the labels; obtaining embedding vectors of radar platforms and embedding vectors of targets based on the one-hot vectors; obtaining the joint embedding vectors based on the embedding vectors of the radar platforms and the embedding vectors of the targets.

3. The method according to claim 1, wherein The method for obtaining a weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using a GCN network comprises the following steps: obtaining an initial weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using the GCN network; obtaining an initial score vector based on the initial weight matrix and the sample deep residual signal features, and obtaining an initial identification result based on the initial score vector; training the GCN network based on the initial identification result and a true result by using a combined loss function until the loss function value meets a set requirement, and obtaining the weight matrix based on the efficient features and the symmetric normalized adjacency matrix by using the trained GCN network.

4. The method of claim 1, wherein, The method for obtaining efficient features based on the sample deep residual signal features and the joint embedding vectors comprises the following steps: performing feature splicing on the sample deep residual signal features and the joint embedding vectors by using attention weighting to obtain the efficient features.

5. The method of claim 1, wherein, The method for obtaining a score vector based on the deep residual signal features and the classifier comprises the following steps: performing convolution fusion on the deep residual signal features and the classifier to obtain the score vector.

6. The method according to claim 3, wherein The combined loss function is represented as: ； In the formula, This represents the loss value of the initial recognition result. Indicates the number of radar platforms. Indicates the number of targets. Indicates the first in the true result The actual value of each element Indicates the initial recognition result for the first The probability that an element is predicted to be 1. This indicates the composite weight.

7. A multi-platform multi-tag radiation source individual identification system, characterized in that, The method comprises the following steps: a data acquisition module is configured to obtain the number data of all radar platforms and the number data of all targets in a set region, sample radar data, and radiation source signal data in the set region; a label dependency modeling module is configured to construct labels based on the number data of the radar platforms and the number data of the targets, and obtain joint embedding vectors based on the labels, and construct a symmetric normalized adjacency matrix based on the labels; The feature representation learning module is configured to obtain a sample deep residual signal feature based on the sample radar data by using a deep residual network, and configured to obtain an efficient feature based on the sample deep residual signal feature and the joint embedding vector, and configured to obtain a deep residual signal feature based on the emitter signal data by using the deep residual network; The GCN classifier learning module is configured to obtain a weight matrix based on the efficient feature and the symmetric normalized adjacency matrix by using a GCN network, and configured to take the weight matrix as a classifier; The data output module is configured to obtain a score vector based on the deep residual signal feature and the classifier, and configured to obtain a radar platform number and a target number of the emitter signal data based on the score vector.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-platform multi-label emitter individual identification method in any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the multi-platform multi-label emitter individual identification method in any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the multi-platform multi-label emitter individual identification method in any one of claims 1-6.

Images