Multi-platform multi-label radiation source individual identification method, system, equipment, medium and product

Abstract

The invention discloses a multi-platform multi-label radiation source individual identification method, system and device, a medium and a product, and relates to the technical field of target identification, and the method comprises the steps: building a label based on a radar platform and the number data of a target; obtaining a joint embedding vector based on the label and constructing a symmetric normalized adjacent matrix; acquiring sample radar data; using a deep residual network to obtain sample depth residual signal features based on the sample radar data; efficient features are obtained based on the sample depth residual signal features and the joint embedding vector; adopting a GCN network to obtain a classifier based on the efficient features and the symmetric normalized adjacency matrix; radiation source signal data in a set area are acquired, and depth residual signal features are obtained by using a depth residual network; obtaining a score vector based on the deep residual signal features and a classifier; and obtaining a radar platform number and a target number of the radiation source signal data based on the fractional vector. According to the invention, the identification speed and accuracy of the radiation source signal data can be improved.

Description

Technical Field

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

[0002] The individual characteristics of radar emitters are formed by parasitic modulation generated by key transmitter components (such as the transmitting tube and power amplifier). These characteristics cannot be avoided, eliminated, or forged. Differences in modulation form and modulation amount make them unique, making individual identification possible. Specific Emitter Identification (SEI) uniquely identifies radio emitters by measuring the external characteristics of intercepted electromagnetic signals. This technology is widely used in radar target detection and perception.

[0003] Existing SEI methods fall into two main categories: those based on manual feature extraction and those based on deep learning. Manual feature extraction methods extract signal features from the time, frequency, or time-frequency domains, and combine them with classifiers such as support vector machines (SVMs) and decision trees to complete classification tasks. However, these methods rely heavily on expert knowledge and have poor adaptability to complex electromagnetic environments. In contrast, deep learning methods automatically learn signal features to identify individual radiation sources, outperforming manual feature extraction methods. Common deep learning models include convolutional neural networks (CNNs), residual networks (ResNets), complex-valued neural networks, gated recurrent units (GRUs), and transformer models.

[0004] However, the two aforementioned SEI methods primarily rely on a single radar platform to classify and identify multiple targets. Due to limited data sources, single-platform classification and identification may not fully cover the target area and is easily affected by local environmental factors (such as terrain obstruction and electromagnetic interference). This results in reduced classification and identification accuracy, poor robustness, and a perception range that cannot meet actual requirements. Therefore, multi-platform distributed fusion recognition technology has emerged. Currently, multi-platform distributed identification of individual emitters faces heterogeneous feature distribution differences, such as complex land-based electromagnetic interference, signal time-varying caused by dynamic sea conditions, and airborne Doppler frequency shift and limited onboard processing. Traditional homogeneous fusion methods face difficulties in effectively training and achieving convergence due to multi-platform feature mismatch, resulting in limited recognition performance. Therefore, how to utilize multi-platform distributed fusion to identify individual emitters in heterogeneous feature scenarios is an urgent problem that needs to be solved.

[0005] Currently, there are several approaches to addressing this issue: multi-label classification methods; receiver coloration compensation methods; and transmitter fingerprint mapping modeling methods. Because receiver coloration removal requires complete multi-channel full-pulse data for training and debugging, and the data collected in actual measurement environments is limited, the applicability of this method is limited. Furthermore, due to the nonlinear time-varying characteristics of radar signals in complex and diverse application scenarios, constructing a transmitter fingerprint mapping function is difficult to use as an accurate mathematical model to effectively characterize the differences between individual radiators under multi-scenario migration conditions, resulting in certain limitations for this approach. In contrast, multi-label classification methods achieve feature decoupling by constructing a multidimensional label system. This approach is independent of the integrity of multi-channel data and the consistency of signal distribution, and offers significant advantages in SEI problems with large heterogeneous feature distribution differences.

[0006] Multi-label feature recognition uses deep learning technology to extract effective features from signal features and perform multi-label classification to achieve end-to-end recognition. In recent years, some existing studies have made a series of progress in multi-label feature recognition around the three dimensions of feature representation optimization, model lightweighting, and cross-domain adaptability. However, these existing studies still have fundamental limitations: based on the single-platform multi-label classification method, the "platform-target" combination is modeled as an independent discrete label, resulting in the label space growing exponentially with the number of platforms, and a separate feature extractor needs to be designed for each label. This not only ignores the correlation and dependency between the platform and the target, but also introduces label noise interference due to the combination of unrelated and weakly related labels, resulting in a waste of computing resources and a decline in classification performance. It is also unable to adapt to the label space expansion needs of multi-platform systematic recognition. Summary of the Invention

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

[0008] To achieve the above objectives, this application provides the following solutions: In a first aspect, the present application provides a multi-platform multi-label radiation source individual identification method, comprising: Obtain the serial number data of all radar platforms and all targets in the set area; Constructing a label based on the number data of the radar platform and the number data of the target; obtaining a joint embedding vector based on the label; and constructing a symmetric normalized adjacency matrix based on the label; Acquire radar data of all the targets received by all the radar platforms in a set area as sample radar data; Using a deep residual network, a sample depth residual signal feature is obtained based on the sample radar data; Obtaining efficient features based on the sample depth residual signal features and the joint embedding vector; A GCN network is used to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix, and the weight matrix is ​​used as a classifier; Acquiring radiation source signal data within a set area; the radiation source is a target within the set area; The deep residual network is used to obtain a deep residual signal feature based on the radiation source signal data; a score vector is obtained based on the deep residual signal feature and the classifier; and a radar platform number and a target number of the radiation source signal data are obtained based on the score vector.

[0009] In a second aspect, the present application provides a multi-platform multi-label radiation source individual identification system, comprising: The data acquisition module is used to obtain the number data of all radar platforms and all targets in the set area, sample radar data, and radiation source signal data in the set area; a label dependency modeling module, configured to construct labels based on the numbering data of the radar platform and the numbering data of the target, obtain a joint embedding vector based on the labels, and construct a symmetric normalized adjacency matrix based on the labels; a feature representation learning module, configured to employ a deep residual network to obtain sample depth residual signal features based on the sample radar data, further configured to obtain efficient features based on the sample depth residual signal features and the joint embedding vector, and further configured to employ the deep residual network to obtain depth residual signal features based on the radiation source signal data; A GCN classifier learning module is used to use the GCN network to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix, and use the weight matrix as a classifier; A data output module is used to obtain a score vector based on the depth residual signal feature and the classifier, and to obtain a radar platform number and a target number of the radiation source signal data based on the score vector.

[0010] In a third aspect, the present application provides 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 steps of the multi-platform multi-label radiation source individual identification method described above.

[0011] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the multi-platform multi-label radiation source individual identification methods described above.

[0012] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the multi-platform multi-label radiation source individual identification methods described above.

[0013] According to the specific embodiments provided in this application, this application has the following technical effects: The present application provides a multi-platform multi-label radiation source individual identification method, system, equipment, medium and product, which constructs labels based on the numbering data of all radar platforms and the numbering data of all targets in a set area, and constructs a joint embedding vector and a symmetric normalized adjacency matrix between the radar platform and the target based on the label, retaining only the effective targets that can be detected by each radar platform, eliminating the influence of irrelevant label combinations on the individual identification of the radiation source, and can improve the correlation dependency between the radar platform and the target, improve the computing resource utilization and recognition performance. A deep residual network is used to obtain sample deep residual signal features based on sample radar data to realize the extraction of deep signal features, and to obtain efficient features based on the deep residual signal features and the joint embedding vector. A GCN network is used to process the efficient features and the symmetric normalized adjacency matrix to obtain a classifier. The classifier is trained by combining the loss function of the GCN network, and is obtained by using the trained GCN network based on the efficient features and the symmetric normalized adjacency matrix. It can solve the long-tail distribution problem of labels and improve the recognition robustness of small sample categories, thereby meeting the label space expansion requirements for systematic identification covering multiple radar platforms. In practical applications, a deep residual network is used to obtain deep residual signal features based on the radiation source signal data within the set area, and a score vector is obtained based on the deep residual signal features and the classifier, and then the radar platform number and target number of the radiation source signal data are obtained, thereby improving the recognition speed and accuracy of the radiation source signal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0015] Figure 1 This is a flowchart of a multi-platform multi-label radiation source individual identification method in one embodiment of the present application; Figure 2 A framework diagram of a multi-platform, multi-label radiation source individual identification method provided in one embodiment of the present application; Figure 3 This is an overall flow chart of the multi-platform multi-label radiation source individual identification method provided in one embodiment of the present application; Figure 4 This is an overall architecture diagram of the feature representation part provided in one embodiment of the present application; Figure 5 A schematic diagram of the structure of a multi-platform multi-label radiation source individual identification system provided in one embodiment of the present application; Figure 6 A schematic diagram of the relative layout positions of a radar platform and a target provided in another embodiment of the present application; Figure 7 A schematic diagram of the combined loss function parameter settings and optimization results provided in another embodiment of the present application; Figure 8 A comparison chart showing the impact of different GCN layer numbers on recognition accuracy provided in another embodiment of this application; Figure 9 A comparison chart showing the impact of different dimensional transformation strategies on recognition accuracy provided in another embodiment of the present application; Figure 10 A comparison chart of recognition accuracy under three loss functions provided in another embodiment of the present application; Figure 11 A schematic diagram of the confusion matrix of a multi-platform multi-label radiation source individual identification method and a traditional CNN multi-classification network model provided in another embodiment of the present application; Figure 12 A schematic diagram of performance comparison results provided by another embodiment of the present application; Figure 13 A schematic diagram of a biased test confusion matrix provided in another embodiment of the present application. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0017] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0018] In an exemplary embodiment, Figure 1 、 Figure 2 and Figure 3 As shown, a multi-platform multi-label radiation source individual identification method is provided, comprising: Step 100: Obtain the serial number data of all radar platforms and all targets within a set area, and construct labels based on the serial number data of the radar platforms and the targets.

[0019] Step 200: Obtain a joint embedding vector based on the labels and construct a symmetric normalized adjacency matrix based on the labels.

[0020] Step 300: Acquire radar data of all targets received by all radar platforms within a set area as sample radar data.

[0021] In step 400, a deep residual network is used to obtain sample depth residual signal features based on sample radar data, and an efficient feature is obtained based on the sample depth residual signal features and the joint embedding vector.

[0022] Step 500: A GCN network is used to obtain a weight matrix based on efficient features and a symmetric normalized adjacency matrix, and the weight matrix is ​​used as a classifier.

[0023] Step 600: Acquire emitter signal data within a set area. The emitter is a target within the set area. A deep residual network is used to obtain deep residual signal features based on the emitter signal data. A score vector is generated based on the deep residual signal features and a classifier. Based on the score vector, the radar platform number and target number of the emitter signal data are obtained.

[0024] Among them, the radar platform refers to a radar signal receiver platform that can receive radar data (i.e. radar signal data) of the target (i.e. radiation source).

[0025] As an optional implementation, to improve the performance of individual emitter identification, the process of obtaining a joint embedding vector based on the label in step 200 includes: obtaining a one-hot vector based on the label; obtaining an embedding vector for the radar platform and an embedding vector for the target based on the one-hot vector; and obtaining a joint embedding vector based on the embedding vectors of the radar platform and the target.

[0026] For example, in actual radar detection applications, radar signal data usually comes from multiple radar platforms and contains radar data of different targets. Implicit modeling can enhance the expressiveness of labels and at the same time serve as a splicing of label features and data features, making it easier for the classifier to fuse label features and signal features to improve classification and recognition performance.

[0027] In the implicit modeling of label dependency, considering the radar signal data from any radar signal receiver platform and for any target, the label of the "platform-target" combination is defined based on the acquired radar platform number data and the target number data, which is denoted as ,in and They represent the radar platform number and the target number respectively, and are used to indicate the source platform and corresponding target of the radar data. The unique hot vector obtained based on the label is shown in formula (1).

[0028] (1) Where, represents a one-hot vector, and Represent the unique hot encoding of the radar platform and the target respectively, . and Respectively represent the dimensions and The real vector space of . and are the total number of radar platforms and the total number of targets respectively.

[0029] In order to improve data operation efficiency while retaining key information that is easy to extract, 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. 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).

[0030] (2) (3) Where, 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, is the embedding dimension, and Respectively represent the dimensions and The real matrix space of . represent the radar platform embedding vector, the target embedding vector and the joint embedding vector respectively.

[0031] In the explicit modeling of label dependencies, the complex dependency structure between labels is captured by constructing the association relationship between labels, and a correlation matrix based on co-occurrence patterns is designed to guide the information propagation between nodes in the GCN network. In order to ensure numerical stability and explicitly model sub-links, the co-occurrence matrix is ​​defined as As shown in formula (4).

[0032] (4) in, Represents the co-occurrence matrix Elements in , used to quantize labels and tags The strength of the association between Representation Label and The statistical frequency (or similarity) of co-occurrence in the data. is an optional small positive number used to prevent the diagonal elements of the co-occurrence matrix from being zero. represents the Kronecker delta function, ensuring that each label and The self-connection weight is at least , thus avoiding numerical instability or graph structure breakage caused by zero co-occurrence frequency, Satisfies formula (5): (5) In order to build label dependencies by explicitly assigning self-connection and neighbor weight ratios, a weighted adjacency matrix is ​​constructed , as shown in formula (6).

[0033] (6) in, Represents the weighted adjacency matrix Middle Rank Elements of the column used to quantize labels and tags The connection weight between . It is the preset self-connection weight, usually set to 0.8, which indicates the dependency of the label node on its own features. is an optional small regularization term to ensure numerical stability. The weighted adjacency matrix The weights of the off-diagonal elements are distributed in proportion to their co-occurrence frequencies, and the total neighbor weight is , usually set to 0.2, which implies the strength of the dependency between labels while balancing the problems of overfitting and oversmoothing. Represents the co-occurrence matrix Middle Rank Elements of the column used to quantize labels and tags The strength of the association between them.

[0034] In order to construct the symmetric normalized adjacency matrix , for any node , should satisfy formula (7): (7) Where, Represents the weighted adjacency matrix After symmetric normalization Rank Elements of the column used to quantize labels and tags The normalized connection weights between . The diagonal elements of the correspondence matrix of the symmetric normalized adjacency matrix represent the labels The normalized degree is used to characterize the total connection strength of labels in the graph structure and supports symmetric normalization operations. represents the augmented matrix No. Rank The elements of the column are represented by and the identity matrix The elements at corresponding positions are added together to explicitly retain the self-connection weights, ensure the integrity of the graph structure, and serve as an intermediate transition for the symmetric normalization operation. Indicates the dimension The identity matrix Middle Rank Elements of the columns used to construct the augmented matrix , the elements on the diagonal of the unit matrix are 1, and the rest are 0, which is used to force each label to add self-connection to avoid abnormal graph convolution calculation due to no self-connection of labels, while strengthening the retention of the label's own features.

[0035] In terms of label dependency modeling, an implicit modeling module with multi-level label embedding establishes a unified semantic feature representation for heterogeneous labels. Discrete platform and target numbers are mapped into continuous label vectors with certain correlations, which are defined as label features. Then, a sparse correlation matrix is ​​constructed through explicit modeling, and a matrix pruning strategy is implemented to construct a symmetric normalized adjacency matrix, effectively eliminating redundant label interference. A reweighting mechanism is also introduced to adjust the weights of the graph structure, maintaining a precise representation of label correlations while suppressing oversmoothing during graph convolution, laying the foundation for subsequent cross-modal feature interaction.

[0036] As an optional implementation, to facilitate subsequent feature extraction, the sample radar data within the set area obtained in step 300 and the emitter signal data obtained in step 600 are both preprocessed. The preprocessing process specifically includes: performing a Fourier transform on the radar signal data to obtain complex frequency domain components of the radar signal data; obtaining mapping features based on the amplitude spectrum of the complex frequency domain components; and performing linear scaling on the mapping features to obtain preprocessed radar signal data (e.g., sample radar data, emitter signal data).

[0037] For example, in scenarios where a fixed receiver (i.e., a radar platform) works in conjunction with a single radiation source (i.e., a target), the signal modeling process can be implemented through nonlinear device characteristic analysis. Under this time-separated reception mechanism, the need to separate the signals of multiple radiation sources is circumvented, as their physical layer characteristics are primarily derived from the nonlinear distortion effects of the power amplifier (PA). Assume that the RF modulated signal input to the PA is represented by formula (8): (8) in, It represents the time domain sampling value of the RF signal after the baseband modulation signal is modulated by the carrier, in represents the input after modulation, Characterize the baseband modulation waveform, and Corresponding to the carrier frequency and sampling rate parameters respectively. It is the time sampling index, which is convenient for processing when converting continuous signals into discrete signals.

[0038] In order to characterize the non-transmission characteristics of the power amplifier, a polynomial series expansion model is adopted. The output signal of the power amplifier in the single-hop communication scenario can be expressed as formula (9): (9) Where, Indicates the The power amplifier output signal of a radiation source, For the The polynomial expansion coefficients (i.e., series expansion coefficients) of the power amplifiers of different radiation sources. Different coefficients represent the differences in power amplifiers between different radiation sources. K represents the nonlinear order. Indicates the number of radiation sources.

[0039] The receiver (i.e. radar platform) obtains the discrete sampling signals from a radiation source Satisfying formula (10): (10) in, Indicates the The channel fading factor from the transmitter of the radiation source to the receiver, For the The channel additive noise term of the radiation source, is the sampling block index of the received signal.

[0040] Further expanding formula (10) yields the analytical formula for the received signal, as shown in formula (11).

[0041] (11) The above steps quantify the amplifier distortion effects through a nonlinear series expansion, while explicitly modeling channel attenuation and noise interference. This converts heterogeneous radiation source signals into dimensionally consistent and comparable feature vectors, facilitating consistent representation across samples. To facilitate subsequent feature extraction, further signal domain transformation and statistical normalization are required.

[0042] For time domain signals Perform Fourier transform at point W to obtain the frequency domain representation, as shown in formula (12).

[0043] (12) in, For the The complex frequency domain components of a radiation source, usually its amplitude spectrum As the analysis object. is the frequency domain sampling index, indicating the frequency points. Is an imaginary unit.

[0044] In order to suppress the difference in spectral dynamic range, logarithmic transformation and offset compensation are applied to the amplitude spectrum of the complex frequency domain component, as shown in formula (13).

[0045] (13) in, , which is used to avoid numerical singularities caused by zero-value input and enhance the discrimination of low-energy frequency bands. Indicates the The amplitude spectrum characteristics of the complex frequency domain components of the radiation source after logarithmic transformation and offset compensation, The frequency domain sampling index.

[0046] Based on the statistical characteristics of all samples (i.e., sampled signals), the features are mapped to the standardized interval, as shown in formula (14).

[0047] (14) in, Indicates the The amplitude spectrum characteristics of the radiation source after normalization, and Represent the mean and variance of all samples, is a numerical stability constant.

[0048] Finally, through linear scaling Constrained to the range (0,1), as shown in formula (15).

[0049] (15) in, Indicates the The amplitude spectrum characteristics of the radiation source after linear scaling, Represent the maximum and minimum values ​​of the amplitude spectrum characteristics of all radiation sources respectively.

[0050] As an optional implementation, in step 400, efficient features are obtained based on the sample depth residual signal features and the joint embedding vector, including: using attention weighting to perform feature splicing on the sample depth residual signal features and the joint embedding vector to obtain efficient features.

[0051] For example, the sample radar data is first upgraded in dimension through signal reconstruction, and a deep residual network architecture is used to stack jump connection modules to simultaneously capture local detail features and global feature information of the signal through multi-level residual learning. Based on the joint embedding vector and the sample deep residual signal features extracted by the deep residual network, the two are fused at the splicing layer, and feature dimension selection is implemented through a dual-channel attention weighting module, which applies specific weights to the feature channels of the signal domain and the label domain respectively to achieve efficient feature representation. The overall architecture of the feature representation part is as follows: Figure 4 shown.

[0052] Using ResNet as the base network model and the deep structure of the ResNet-101 network, we extract signal features from the dimensionally transformed sample radar data through convolutional layers and residual blocks. We then perform average pooling and dimension transformation on the extracted signal features to obtain the sample depth residual signal features. We then use attention weighting to concatenate the sample depth residual signal features and the joint embedding vector to obtain efficient features, as shown in Formula (16).

[0053] (16) Where, represents sample radar data, represents efficient features, Represents the sample depth residual signal characteristics, Indicates the attention weighting of the features after the sample depth residual signal features and the joint embedding vector features are fused together. Indicates dimension conversion processing, Indicates that CNN convolutional neural network is used to extract features. represents the average pooling operation, represents the attention weighting mechanism, represents the dimension of efficient features, express dimensional efficient feature vector space.

[0054] As an optional embodiment, to ensure recognition accuracy, step 500 includes: using a GCN network to obtain an initial classifier based on efficient features and a symmetric normalized adjacency matrix. Based on the initial classifier and the sample depth residual signal features, an initial score vector is obtained, and an initial recognition result is obtained based on the initial score vector. Using a combined loss function, the GCN network is trained based on the initial recognition result and the actual result until the loss function value meets the set requirements. The trained GCN network is then used to obtain a classifier based on the efficient features and the symmetric normalized adjacency matrix.

[0055] The combined loss function is shown in formula (17).

[0056] (17) Where, represents the loss value of the initial recognition result, represents the number of radar platforms, represents the number of targets, Indicates the actual result The true value of the element, Indicates the initial recognition result for The probability that an element is predicted to be 1, Represents a composite weight.

[0057] For example, due to the differences in detection and stealth performance among various radar platforms and targets, there is a certain combination relationship between each radar platform and the detected target. If the multi-label classification problem is simply simplified into a binary classification problem mathematically, the correlation structure between the radar platform and the target will often be ignored, so that the problem analysis only stays at the level of formulaic structural reasoning. When the number of platform and target categories is large, there will be high computational complexity and difficulty in expansion, which is essentially limited. The idea of ​​recognition and classification of multiple targets in the sample based on the probabilistic graphical model in the GCN network is applied to the problem of one-dimensional radar signal feature extraction, and the label correlation between the radar platform category and the target category is explicitly modeled, which can greatly save computing resources and achieve fast analysis, and effectively avoid the influence of unrelated or weakly related pairs between "platform-target" on the accuracy of predicted categories. In the GCN network, node features are represented by the adjacency matrix To spread information, each node The feature update of is shown in formula (18).

[0058] (18) in, Representation Label After the The feature representation after layer GCN network processing, For the The weight matrix of the learnable parameters of the layer GCN network, is the activation function, the neighbor term Aggregate neighbor information by weighting co-occurrence frequency.

[0059] go through The result obtained after layer GCN network processing is as shown in formula (19).

[0060] (19) in, The feature representation initially input into the first layer of the GCN network (i.e., efficient feature ), For passing The output of the layer GCN network, Indicates the The bias term of the GCN network is used to adjust the output of the network and increase the expressive power of the network model.

[0061] An 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 characteristics. 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 true result using the combined loss function shown in formula (17) until the loss function value meets the set requirements. The classifier is then obtained using the trained GCN network based on the efficient features and the symmetric normalized adjacency matrix.

[0062] For example, in multi-label classification tasks, especially the problem of multi-platform, multi-label radiation source individual identification, class imbalance and modeling inter-label dependencies are two key issues. Traditional loss functions (such as cross-entropy loss) perform poorly when dealing with class-imbalanced data and cannot explicitly model inter-label dependencies. Therefore, a combined loss function (Combined Loss, CB) is designed, which integrates focal loss (FL), multi-label loss (ML), and class imbalance loss (IC). By dynamically adjusting 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 model's classification performance on multi-label data is improved.

[0063] Since multi-label classification is involved, the multi-label loss function MLL is selected as the basis function, as shown in formula (20).

[0064] (20) in, Represents the multi-label loss function value, , .

[0065] In order to achieve the balance of weights of different categories through dynamic adjustment, suppress the gradient contribution of simple samples (high confidence), and focus on the learning of difficult samples (low confidence), the FL weight is introduced. , Inverse Frequency (IF) weight , and dynamic index adjustments , as shown in formula (21).

[0066] (twenty one) Among them, when Near 1 o'clock, Approaching 0, reducing the loss contribution of simple samples; when Near 0 o'clock, Increase and strengthen the learning of difficult samples. The real result is The number of times the element is not 0. 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; is a smoothing term.

[0067] Combining the three in formula (21) gives the composite weight , as shown in formula (22).

[0068] (twenty two) The composite weight The loss term only acts on positive samples, and the negative samples keep the original weights to form the final combined loss function. Formula (17) is combined with Formula (22) to obtain Formula (23).

[0069] (twenty three) As an optional implementation, in order to improve classification performance, the process of obtaining a score vector based on the depth residual signal feature and the classifier in step 600 includes: performing convolution fusion on the depth residual signal feature and the classifier to obtain a score vector.

[0070] For example, the deep residual signal features are convolved with the graph structure features (i.e., the classifier, which is a weight matrix containing several parameter weights) output by the trained GCN network to obtain the score vector , as shown in formula (24).

[0071] (twenty four) in, Represents the result of the last layer of GCN operation (i.e., classifier), Represents the depth residual signal characteristics.

[0072] In addition, in order to ensure that the combined loss function supports end-to-end optimization, it is necessary to verify the derivability of the combined loss function with respect to the parameters. Verify that the gradients of the positive sample part and the negative sample part of the combined loss function are both differentiable.

[0073] For positive samples ( ) and negative samples ( ), as shown in formulas (25) and (26), the derivative of the loss function for positive and negative samples can be obtained: (25) (26) Then the total gradient of the combined loss function is the weighted sum of the gradients of positive samples and negative samples as shown in formula (27).

[0074] (27) From the above formula, we can see that the hyperparameter in the combined loss function is a logarithmic function and exist Internally differentiable; exponential function Infinitely differentiable in any real number field; polynomials and power functions and exist Internally differentiable; the product and linear combination of all derivative terms remain differentiable, so the loss function right The gradient exists everywhere and is continuous, meeting the requirements of the back propagation.

[0075] Based on the same inventive concept, the present application also provides a multi-platform, multi-label radiation source individual identification system. The implementation solution provided by this system is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations in the following multi-platform, multi-label radiation source individual identification system embodiment 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.

[0076] In an exemplary embodiment, Figure 5 As shown, a multi-platform multi-label radiation source individual identification system is provided, comprising: The data acquisition module is used to obtain the number data of all radar platforms and all targets in the set area, sample radar data and radiation source signal data in the set area.

[0077] The label dependency modeling module is used to construct labels based on the numbered data of the radar platform and the numbered data of the target, obtain a joint embedding vector based on the labels, and construct a symmetric normalized adjacency matrix based on the labels.

[0078] The feature representation learning module is used to obtain sample depth residual signal features based on sample radar data using a deep residual network, and is also used to obtain efficient features based on sample depth residual signal features and a joint embedding vector. It is also used to obtain deep residual signal features based on radiation source signal data using a deep residual network.

[0079] The GCN classifier learning module is used to obtain a weight matrix based on efficient features and a symmetric normalized adjacency matrix using the GCN network, and use the weight matrix as a classifier.

[0080] The data output module is used to obtain a score vector based on the depth residual signal feature and the classifier, and obtain the radar platform number and target number of the radiation source signal data based on the score vector.

[0081] In one exemplary embodiment, combining the multi-platform, multi-tag emitter identification method described in the previous examples, experimental analysis of the identification method was conducted using measured data from two radar signal receiver platforms (Platform 1 and Platform 2) on six targets (Targets 1 to 6). The target emitter signal data was collected from the same model of X-band maritime navigation radar, using LFM modulation. The radar parameters are shown in Table 1. The radar platform's receiver used a portable ultra-wideband real-time spectrum analyzer with an instantaneous receiving bandwidth of 80 MHz.

[0082] Table 1 Radar parameters

[0083] To adapt to the multi-platform multi-target recognition task, the different platform-target combinations were divided into 12 experimental groups (numbered 1 to 12), where platform 1 and target i (i=1-6) corresponded to groups 1 to 6, and platform 2 and target i (i=1-6) corresponded to groups 7 to 12. The radar signal receiver platform and the six target emitters were deployed at set positions (corresponding to platform 1 and platform 2, targets 1-target 6). The farthest and closest distances between the two platforms and the six targets were 21.8km and 17.8km, 11km, and 6.3km, respectively. The locations and surrounding areas of each platform and target were not on the waterway, with fewer ships and less environmental interference. The relative positions of the platforms and targets are as follows: Figure 6 The acquisition scenario is as follows: six target radars are powered on simultaneously, and the transmit power of each navigation radar is manually adjusted to make the signal of a single target emitter relatively strong. At this time, the receivers of the two radar platforms identify the target emitter with the strongest received signal.

[0084] Because actual data collection did not include some platform and target combinations, the measured data only includes groups 1-4, 7, and 10-12. To minimize the impact of time on experimental results, all measured data without time offset was 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 measurement served as the time offset test set to assess the model's applicability under time distribution conditions. Based on this, classification tasks were constructed for two scenarios: without time offset (data within one month) and with time offset (data outside one month and within six months). The experimental design split the dataset into an 8:2 ratio, as shown in Table 2.

[0085] Table 2 Dataset division table under two scenarios

[0086] The experiments were conducted using the PyTorch 2.0.1 framework on Windows 10 Professional. The hardware configuration included an Intel(R) Core™ i9-9900K CPU (3.6 GHz) and an NVIDIA GeForce RTX A6000 GPU. Accuracy was used as the evaluation metric. The training epochs were 20, with an initial learning rate of 0.00125. Cosine annealing was used to update the learning rate in each epoch, and a regularization term was introduced to prevent overfitting.

[0087] The design of loss function is crucial for classifier optimization in recognition methods. In the parameter setting of the combined loss function, in order to determine the hyperparameters in the above combined loss function 、 and The optimal value of , this experiment uses a phased grid search method to tune the parameters on an unbiased test set to balance the prediction accuracy in the class imbalance scenario. First, based on the basic value determined in the pre-experiment, the focus loss adjustment factor is fixed. and inverse frequency attenuation coefficient and momentum factor The two parameters in the experiment are searched for five sets of discrete values ​​of the third hyperparameter on the validation set, and the validation results are cross-validated to obtain the local optimal value of the third hyperparameter under the experimental conditions. The specific hyperparameter settings and results are shown in Figure 7 shown. Figure 7 Part (a), part (b) and part (c) are Time The results of the search for the best Time The results of the optimization and Time Optimization results. Figure 7 The results shown show that under the current data and experimental conditions, when When , the accuracy of the model on the unbiased test set increased to 95% in 5 epoch training cycles and gradually approached convergence, proving that the model achieved the best performance when this parameter was set, and the optimization process was stable.

[0088] Based on the theory of message passing mechanism, the number of layers of the GCN network determines the aggregation range of neighborhood information of node features, which directly affects the model's ability to capture graph structure information, and thus affects the model's recognition and classification capabilities. However, excessive stacking of layers can easily lead to oversmoothing, causing node features to tend to be homogenized and resulting in blurred classification boundaries. In the structural selection of the GCN network, in order to quantitatively analyze the impact of the number of GCN layers, a comparative architecture with model layers ranging from 1 to 5 GCN layers was constructed, and the same initialization strategy was used to control variables. The model performance was evaluated by observing the classification accuracy. The results are shown in the figure. Figure 8 shown.

[0089] Results show that when the number of GCN layers is 2, the model achieves over 95% classification accuracy at just epoch 5, while other GCN layers fail to reach 90% classification accuracy at this point. At the same training epoch, the model achieves high classification accuracy while also exhibiting rapid convergence, ultimately reaching a peak accuracy of 97%. These experimental results validate the effectiveness of the model's structural design. They also demonstrate that deep GCNs exhibit a feature degradation effect, whereby some features are over-smoothed as the number of convolutional layers increases. The two-layer GCN architecture achieves an optimal balance between aggregating local neighborhood information and preserving global structure, providing a theoretical basis for subsequent model design.

[0090] In terms of the choice of dimensionality transformation, the dimensionality transformation of the input data (the splicing feature of each sample is 1×512 dimensions) is crucial to the model's feature extraction capability. To address the feature adaptation problem of one-dimensional radar signals in a two-dimensional convolutional architecture, based on the local perception field theory of convolutional neural networks, five groups of data dimensionality transformations are designed, namely (1, 512), (2, 256), (4, 128), (8, 64), and (16, 32). The classification accuracy of different groups is compared to select a suitable dimensionality reorganization scheme for the model. At the same time, appropriate training truncation can be performed based on its convergence speed to save computing resources. The results are shown in the figure. Figure 9 shown.

[0091] The results show that each group can achieve a recognition accuracy of more than 97% after 20 sets of training, but the 16×32 group strategy can achieve a classification accuracy of more than 95% in 5 epochs, and its convergence speed is faster than other groups. The reason is that this dimension has a better geometric adaptability to the residual convolutional feature extraction network. Through parameter sharing efficiency optimization and computational graph fusion, while maintaining a certain feature expression capability, the overhead of computing resources is reduced, which can achieve rapid convergence.

[0092] Through the above comparative experimental parameter setting optimization experiment, it can be seen that the selection , the number of GCN layers is 2, and the dimension transformation strategy is (16, 32) to achieve the local optimal performance of the model. At the same time, because the model has a faster convergence ability under this setting, the number of training rounds can be set to 10 according to the computer computing power and the requirements for real-time operation.

[0093] 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 the focus function (FL). The recognition accuracy of the recognition method is evaluated by training the GCN network on the dataset shown in Table 1 under the above-selected local optimal parameter settings. The results are shown in Figure 1. Figure 10 Experimental results show that the propagation optimization of the designed combined loss function achieves higher recognition accuracy and shorter training time to achieve a certain accuracy compared to the other two methods. This demonstrates the effectiveness of the combined loss function's implicit label modeling and weight adjustment strategy in multi-platform and multi-label scenarios, and is also effective in addressing the problem of class imbalance.

[0094] In order to verify the advantages of graph structure classification models in modeling complex relationships, this experiment compares the performance of the multi-platform multi-label radiation source individual recognition method with the traditional CNN multi-classification network model to evaluate the differences in accuracy and time complexity. The traditional method is based on the CNN classification strategy mechanism, which regards the "platform-target" combination as 12 independent labels. After the ResNet-101 feature extraction network module structure, a high-dimensional fully connected classifier is constructed, and the fully connected layer is used to achieve terminal classification, while ignoring the correlation between the platform and the target. The confusion matrix of the recognition method provided by this application and the CNN-based multi-classification model, as well as the operation time of each training round, are shown as follows: Figure 11 As shown in parts (a) and (b) of the table 3, the and They represent the computation time of the recognition method in this application and the traditional multi-classification model respectively.

[0095] Table 3 Comparison of the computational time complexity of each training round of the two methods

[0096] Depend on Figure 11 From part (a) of the paper, we can see that the overall recognition accuracy of the recognition method in this application can reach more than 97% and the recognition accuracy of each imbalanced category is more than 94%; Figure 11 As shown in part (b), the traditional multi-classification model has an overall recognition accuracy of only 80%, with individual category recognition accuracy just above 74%. This performance is poor for this problem, and there is a 3% probability of misclassification of the three unrelated groups 5, 8, and 9. Experimental results show that the traditional multi-classification model ignores the constraints between platforms and targets (for example, platform 1 can only detect specific targets), resulting in a lack of explicit modeling of their correlations and the introduction of a certain amount of label noise. Furthermore, the independent classification assumption of the fully connected layer ignores the hierarchical dependencies between platforms and targets, failing to capture cross-label collaborative features. This results in invalid label combinations being involved in the calculation and leading to misclassification, which has an unpredictable impact on the final prediction results.

[0097] As shown in Table 3, the GCN-based graph structure model (i.e., classifier) ​​uses an 8-node (2 platforms + 6 targets) bipartite graph design and uses the adjacency matrix to encode the association strength, with an average training time of 154.63 seconds per round. The CNN-based traditional connection design with 12 nodes (2 platforms * 6 targets) takes an average of 170.28 seconds per round. These results indicate that the classification method used in this application under these conditions and this dataset has shorter training time and higher computational efficiency.

[0098] In order to verify the advantages of graph structure classification model (i.e. classifier) ​​in modeling complex relationships, this experiment uses the recognition method provided by this application (i.e. Graph Convolutional Deep Residual Network for Classification (GCDR-Net) that integrates GCN, deep residual network and multi-platform multi-label word embedding) to correspond to Figure 12 The performance of the GCDR-Net in

[15] was compared with that of the simplified recognition method that removed the implicit modeling of labels and the feature splicing part, and the differences in their accuracy in different training rounds and different categories were evaluated. The other parameters were set the same. The comparison results are shown in the figure. Figure 12 shown.

[0099] The experimental results show that the recognition method provided by this application achieves implicit modeling by embedding the "platform-target" combination label into the classifier mapping. Then, by sharing parameters during the training process, the classifiers of all labels can retain the semantic structure. At the same time, the gradients of all classifiers will affect the classifier generation function, so that the average accuracy reaches 95% and 97% in 5 epochs and 20 epochs respectively, and the minimum accuracy of each category is 90% and 92% respectively. However, due to the simplified recognition method that does not introduce label implicit modeling and feature splicing, the average accuracy in 5 epochs and 20 epochs is only 85% and 89% respectively, and the minimum accuracy of each category is 81% and 85% respectively. It can be seen that for this data set and the relevant parameter settings, the introduction of label implicit modeling and feature splicing strategies in this problem has improved the overall classification accuracy and the minimum accuracy of each category by 8% and 7% respectively in 20 training rounds, improving the feature expression ability.

[0100] The method for identifying individual radiation sources needs to have the ability to generalize over time to cope with the feature distribution shift caused by factors such as equipment aging and environmental disturbances. Therefore, the pre-trained model obtained by training with training data within one month and verifying with an unbiased test set is tested on a cross-period test set (the biased test set shown in Table 2) to evaluate the robustness of the model under time drift scenarios. The resulting biased test confusion matrix is ​​shown as follows: Figure 13 shown.

[0101] The experimental results show that the overall accuracy of the recognition method provided in this application can still maintain 89% over a six-month time drift, proving that the proposed scheme has a certain degree of non-uniform stability in the time drift scenario, and some categories can still maintain 100% and 99% recognition accuracy. However, due to the differences in the time drift amount and the amount of training data of each category, the prediction accuracy of individual groups does not reach 80% or even only 65%. This is because in the process of time evolution, the feature hyperplane supported by the limited training samples of some categories fails to fully adapt to the time-varying nonlinearity of the channel, resulting in increased sensitivity of the decision boundary. Therefore, it is necessary to design an updated strategy to improve it.

[0102] Based on the comparative experiments and analysis of the results, this application designs a multi-platform, multi-label, individual radiation source identification method for the multi-platform, multi-label, specific radiation source identification problem. By explicitly constructing a platform-target association matrix and jointly optimizing implicit word embeddings, the model effectively captures the complex dependencies between multiple labels. It implicitly models and fuses label features with signal features, enhancing feature representation. Furthermore, a lightweight GCN network and a residual network are used to achieve hierarchical extraction and fusion of signal features. Experiments demonstrate that the proposed method achieves 97% classification accuracy in complex, field-measured noisy environments, a 17% improvement over traditional multi-classification models without graph structure, while reducing single-round training time by 9.3%. This method also improves performance by 8% over the proposed model without implicit label modeling and feature concatenation, demonstrating both efficiency and robustness. The model maintains 89% recognition accuracy in a six-month time drift scenario, demonstrating its adaptability to feature distribution shifts. However, the performance degradation of individual categories during long-term evolution highlights the shortcomings of the dynamic feature drift compensation mechanism. Future work will explore online incremental learning and adaptive graph structure optimization strategies to further enhance the model's temporal generalization capabilities. This application provides theoretical support and technical path for the intelligent upgrade of multi-platform collaborative perception systems.

[0103] In combination with the description in the above embodiments, this application proposes a multi-platform multi-label radiator individual identification method to address the problems of large differences in heterogeneous feature distribution and poor recognition performance faced by the individual identification of radiators from multiple radar platforms. First, based on the statistical characteristics of radar data, a "platform-target" association matrix is ​​constructed; then, label features are implicitly modeled through word embedding, and the deep residual convolution module is combined to extract deep signal features. At the same time, the data features and label features are attention-weighted spliced; finally, to fully utilize label correlation, the graph convolutional neural network is improved, and the spliced ​​features are used as input to obtain a feature map to generate a classifier. The feature extraction results are fused to generate a classifier to obtain the final classification result. Measured data show that this method achieves dynamic fusion and efficient classification of multi-platform multi-label features. Compared with 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 splicing increases the accuracy by 8%, and the recognition performance fully meets the application requirements of actual scenarios.

[0104] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0105] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0106] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0107] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, 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.

[0108] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A multi-platform multi-label radiation source individual identification method, characterized in that: include: Obtain the serial number data of all radar platforms and all targets in the set area; Building a label based on the number data of the radar platform and the number data of the target; Obtaining a joint embedding vector based on the labels; constructing a symmetric normalized adjacency matrix based on the labels; Acquire radar data of all the targets received by all the radar platforms in a set area as sample radar data; Using a deep residual network, a sample depth residual signal feature is obtained based on the sample radar data; Obtaining efficient features based on the sample depth residual signal features and the joint embedding vector; A GCN network is used to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix, and the weight matrix is ​​used as a classifier; Acquiring radiation source signal data within a set area; the radiation source is a target within the set area; The deep residual network is used to obtain a deep residual signal feature based on the radiation source signal data; a score vector is obtained based on the deep residual signal feature and the classifier; and a radar platform number and a target number corresponding to the radiation source signal data are obtained based on the score vector.

2. The multi-platform multi-label radiation source individual identification method according to claim 1, characterized in that: Obtaining a joint embedding vector based on the labels includes: Obtain a one-hot vector based on the label; Obtaining an embedding vector of the radar platform and an embedding vector of the target 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.

3. The multi-platform multi-label radiation source individual identification method according to claim 1, characterized in that: A GCN network is used to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix, including: Using the GCN network to obtain an initial weight matrix based on the efficient features and the symmetric normalized adjacency matrix; obtaining an initial score vector based on the initial weight matrix and the sample depth residual signal features, and obtaining an initial recognition result based on the initial score vector; The GCN network is trained based on the initial recognition result and the true result using a combined loss function until the loss function value reaches a set requirement, and the weight matrix is ​​obtained based on the efficient features and the symmetric normalized adjacency matrix using the trained GCN network.

4. The multi-platform multi-label radiation source individual identification method according to claim 1, characterized in that: Obtaining efficient features based on the sample depth residual signal features and the joint embedding vector, including: Attention weighting is used to perform feature concatenation on the sample depth residual signal feature and the joint embedding vector to obtain the efficient feature.

5. The multi-platform multi-label radiation source individual identification method according to claim 1, characterized in that: Obtaining a score vector based on the depth residual signal feature and the classifier, including: Convolutionally fuse the depth residual signal feature and the classifier to obtain the score vector.

6. The multi-platform multi-label radiation source individual identification method according to claim 3, characterized in that: The combined loss function is expressed as: ； Where, represents the loss value of the initial recognition result, represents the number of radar platforms, represents the number of targets, Indicates the actual result The true value of the element, Indicates the initial recognition result for The probability that an element is predicted to be 1, Represents a composite weight.

7. A multi-platform multi-label radiation source individual identification system, characterized in that: include: The data acquisition module is used to obtain the number data of all radar platforms and all targets in the set area, sample radar data, and radiation source signal data in the set area; a label dependency modeling module, configured to construct labels based on the numbering data of the radar platform and the numbering data of the target, obtain a joint embedding vector based on the labels, and construct a symmetric normalized adjacency matrix based on the labels; a feature representation learning module, configured to employ a deep residual network to obtain sample depth residual signal features based on the sample radar data, further configured to obtain efficient features based on the sample depth residual signal features and the joint embedding vector, and further configured to employ the deep residual network to obtain depth residual signal features based on the radiation source signal data; A GCN classifier learning module is used to use the GCN network to obtain a weight matrix based on the efficient features and the symmetric normalized adjacency matrix, and use the weight matrix as a classifier; A data output module is used to obtain a score vector based on the depth residual signal feature and the classifier, and to obtain a radar platform number and a target number of the radiation source 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 radiation source individual identification method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the multi-platform multi-label radiation source individual identification method according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the multi-platform multi-label radiation source individual identification method according to any one of claims 1 to 6 is implemented.

Images