A multi-source feature fusion radiation source individual identification method and system

Abstract

The application belongs to the field of wireless communication signal processing and spectrum monitoring, and particularly relates to a multi-source feature fusion radiation source individual identification method and system, the method comprising: synchronously collecting radio frequency signals emitted by a target radiation source; inputting signals collected by each sensor node into a deep neural network respectively, extracting deep feature vectors containing radio frequency fingerprints of the radiation source through multi-layer convolution and pooling operation; adopting discriminant correlation analysis algorithm to fuse and process deep feature vectors from different sensor nodes, obtaining a fused feature vector; the fusion processing maximizes feature correlation between different sensors and maximizes feature difference between different radiation source categories; constructing an integrated classifier model comprising multiple base classifiers, inputting the fused feature vector into the multiple base classifiers, adopting a weighted voting mechanism to make a joint decision, and outputting a category label of the target radiation source. Feature intra-class aggregation and inter-class differentiation are effectively realized.

Description

Technical Field

[0001] This application belongs to the field of wireless communication signal processing and spectrum monitoring technology, specifically relating to a method and system for individual identification of radiation sources through multi-source feature fusion. Background Technology

[0002] With the rapid development of 5G, IoT, and satellite communications, the electromagnetic spectrum is becoming increasingly congested, significantly increasing the difficulty of spectrum regulation. Specific Emitter Identification (SEI) technology uniquely identifies individual radiation sources by analyzing the inherent fingerprint characteristics of transmitter hardware (such as power amplifier nonlinearity, oscillator phase noise, and I / Q modulation imbalance). Existing radiation source identification technologies mainly rely on single-site monitoring. However, in complex electromagnetic environments, single-site received signals are highly susceptible to multipath fading, shadowing effects, and non-line-of-sight transmission, resulting in low signal-to-noise ratios, severely compromised fingerprint characteristics, and difficulty in guaranteeing identification accuracy and system robustness.

[0003] To address the aforementioned issues, multi-sensor data fusion technology has been introduced into this field. Currently, the mainstream fusion methods are mainly divided into data-level fusion, feature-level fusion, and decision-level fusion. While data-level fusion retains the most complete information, it has extremely high requirements for communication bandwidth and cannot eliminate noise interference. Although simple decision-level fusion reduces the amount of data transmitted, it ignores the deep correlation between signals received by different sensors and fails to fully utilize the complementary information in the feature space.

[0004] In summary, existing technologies lack a radiation source identification method that can effectively extract and fuse deep correlation features from multiple sensors in a distributed environment and maximize category differences at the feature level. Summary of the Invention

[0005] This invention provides a method and system for identifying individual radiation sources through multi-source feature fusion, aiming to solve the problems of existing single-station monitoring being greatly affected by channel fading and the low feature utilization rate of traditional fusion methods. This method utilizes the spatial diversity gain obtained from a distributed sensor network, extracts deep semantic features of each node through a deep learning network, introduces a discriminant correlation analysis (DCA) algorithm for feature-level fusion, and finally combines an ensemble learning strategy for joint decision-making, thereby significantly improving the identification accuracy in complex channel environments.

[0006] In a first aspect, the present invention provides a method for individual identification of radiation sources through multi-source feature fusion, comprising the following steps: The radio frequency signals emitted by the target radiation source are collected synchronously by deploying multiple spectrum monitoring sensor nodes within the monitoring area. The signals collected by each sensor node are input into a deep neural network, and deep feature vectors containing the radio frequency fingerprint of the radiation source are extracted through multi-layer convolution and pooling operations. A discriminant correlation analysis algorithm is used to fuse deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. An ensemble classifier model comprising multiple base classifiers is constructed. The fused feature vector is input into multiple base classifiers, and a weighted voting mechanism is used for joint decision-making to output the category label of the target radiation source.

[0007] The deployment of distributed sensor nodes utilizes spatial diversity technology, which can effectively overcome the shortcomings of single-station monitoring in complex electromagnetic environments, such as being greatly affected by multipath fading and shadowing effects. Even if the signal quality of some sensor nodes is poor, the system can still achieve accurate identification by using complementary information from other nodes.

[0008] Existing feature fusion methods often only consider the correlation between features, failing to fully utilize category discrimination information. This invention optimizes both the correlation between sensors and the differences between categories during the fusion process. The former is used to extract common information from multiple perspectives, while the latter enhances discriminative ability. Multiple base classifiers are used for weighted voting; by integrating the advantages of classifiers with different inductive biases, the bias of a single classifier on specific samples is effectively offset, improving the accuracy and stability of recognition.

[0009] As a further limitation of the technical solution of the present invention, a discriminant correlation analysis algorithm is used to fuse deep feature vectors from different sensor nodes, specifically including: The inter-class scatter matrix is ​​calculated for the feature set of each sensor node, and the transformation matrix is ​​obtained through eigenvalue decomposition. The feature set is then subjected to decorrelation transformation. Calculate the inter-set covariance matrix for multiple feature sets after decorrelation transformation, obtain the quadratic transformation matrix through singular value decomposition, and perform correlation maximization transformation on the feature sets. The features of the correlation-maximizing transformation are fused to generate a fused feature vector.

[0010] Specifically, the inter-class scatter matrix is ​​calculated as follows:

[0011] in, The inter-class scatter matrix, This represents the total number of radiation source categories. For the first Number of samples per class For the first Class sample mean, This is the global mean.

[0012] The correlation maximization transformation includes: For the two feature sets after decorrelation transformation and Calculate its inter-set covariance matrix. ;right Perform singular value decomposition Let the quadratic transformation matrix be , Obtain the final transformed features , .

[0013] The fusion of features that maximize correlation transformation includes: Using serial fusion method Alternatively, an additive fusion method may be adopted. .

[0014] By calculating the inter-class scatter matrix and performing eigenvalue decomposition, redundant dimensions that do not contain discriminative information are effectively eliminated, reducing feature dimensionality and computational complexity, while retaining the most discriminative feature components. Singular value decomposition is performed on the decorrelated feature set to maximize the correlation between features from different sensors, enabling the fused features to fully extract common fingerprint information from multi-sensor observation signals and enhancing feature representation capabilities. The cascade fusion method retains all information from each feature set and is suitable for scenarios requiring complete feature information; the additive fusion method maintains the feature dimensionality unchanged and is suitable for scenarios with high computational efficiency requirements. The two fusion methods can be flexibly selected according to actual application needs.

[0015] As a further limitation of the technical solution of this invention, the ensemble classifier model includes three base classifiers: K-nearest neighbors, decision trees, and logistic regression; the fused feature vector is input into the ensemble classifier model, and the ensemble classifier model outputs the category label of the target radiation source, specifically including: The fused feature vector is simultaneously input into each base classifier, and each base classifier outputs a class probability distribution vector; probability distribution vector:

[0016] in, Indicates that under known fusion features And the model is Under the conditions, the target radiation source belongs to the first Confidence level of radiation-like sources.

[0017] A weighted majority voting method is used to jointly decide the outputs of each base classifier to determine the final recognition result.

[0018] The weighted majority voting method is as follows: Weight coefficients are preset based on the accuracy of each base classifier on the validation set. The votes of each base classifier for each category are weighted and summed, and the category with the highest weighted score is selected as the final recognition result. The formula is:

[0019] in, For the final identification result, For indicator functions, For the first A base classifier for each category The vote These are the preset weight coefficients based on the accuracy of each base classifier on the validation set.

[0020] Take the category corresponding to the maximum value. Among the C categories, select the category c that makes the subsequent expression have the largest value.

[0021] K-Nearest Neighbors, decision trees, and logistic regression are based on distance metrics, rule learning, and probabilistic modeling, respectively. These three methods have varying sensitivities to data distribution; their combined use can be complementary, helping to reduce overfitting risk and improve generalization ability. Weighting coefficients are preset based on the accuracy of each base classifier on the validation set, with higher-accurate classifiers receiving greater weight in the decision-making process, making the joint decision result more reliable. Experimental results show that the recognition accuracy of the ensemble classifier is superior to any single base classifier. The base classifiers output class probability distributions instead of hard class labels, preserving uncertainty information in the prediction process and providing richer decision-making basis for subsequent weighted voting.

[0022] As a further limitation of the technical solution of the present invention, after the step of constructing an ensemble classifier model including multiple base classifiers, inputting the fused feature vector into multiple base classifiers, performing joint decision-making using a weighted voting mechanism, and outputting the category label of the target radiation source, the following steps are included: The category labels can be visualized via a display device or reported to the spectrum regulatory center via a communication interface.

[0023] The display device can visualize the category labels for monitoring personnel to view in real time; the communication interface can report the identification results to the spectrum supervision center so as to connect with the existing spectrum monitoring system and realize automated supervision.

[0024] As a further limitation of the technical solution of the present invention, the radiation source radio frequency fingerprint includes one or more hardware fingerprint features among power amplifier nonlinearity, oscillator phase noise, and I / Q modulation imbalance.

[0025] As a further limitation of the technical solution of the present invention, the method constructs a multi-level fusion distributed architecture at the feature level and the decision level: at the feature level, discriminant correlation analysis is used to fuse features from multiple sensors to obtain fingerprint features with strong representation; at the decision level, the classification results are verified a second time by combining the idea of ​​ensemble learning, so that when some sensors are in deep fading or are subject to strong interference, the system can achieve accurate identification through complementary information from other links.

[0026] Secondly, the present invention also provides a multi-source feature fusion radiation source individual identification system, comprising: Multiple distributed spectrum monitoring sensor nodes are deployed within the monitoring area to synchronously acquire radio frequency signals emitted by the target radiation source; The feature extraction module is equipped with a deep neural network, which is used to input the signals collected by each sensor node into the deep neural network and extract deep feature vectors containing the radio frequency fingerprint of the radiation source through multi-layer convolution and pooling operations. The feature fusion module employs a discriminant correlation analysis algorithm to fuse deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. The classification and recognition module is equipped with an ensemble classifier model, which includes multiple base classifiers. The ensemble classifier model is used to input the fused feature vector into the multiple base classifiers, and to make joint decisions using a weighted voting mechanism, and output the category label of the target radiation source. An output module is used to display or report the category labels. The output module includes a display device and a communication interface, used to visualize the category labels or report them to the spectrum monitoring center.

[0027] Distributed sensor nodes provide spatial diversity gain, and the feature extraction module, feature fusion module, and classification and recognition module work together to achieve fully automated processing from signal acquisition to recognition output. The overall system performance is better than the simple superposition of a single module.

[0028] This invention optimizes two objectives simultaneously during feature fusion through discriminant correlation analysis: first, maximizing the correlation between features from different sensors to extract common fingerprint information from multiple perspectives; and second, maximizing the feature differences between different radiation source categories to enhance the discriminative power of the features. This joint optimization of the two objectives results in fused features that contain complementary information from multiple sensors and possess stronger class distinguishability.

[0029] Building upon this foundation, the present invention further constructs an ensemble classifier, utilizing three base classifiers with different inductive biases—K-nearest neighbors, decision trees, and logistic regression—for weighted voting. Since different base classifiers exhibit varying sensitivities to data distribution, the weighted voting mechanism effectively offsets the bias of a single classifier on specific samples, further improving recognition accuracy.

[0030] As can be seen from the above technical solutions, this application has the following advantages: It introduces a discriminant correlation analysis (DCA) mechanism to solve the problem of multi-source feature redundancy, utilizes the correlation of signals observed by different sensors to suppress noise, and explicitly introduces category information during the fusion process. It minimizes intra-class distance while maximizing inter-class distance in the feature space, fundamentally improving the discriminative power of the features. A multi-level fusion distributed architecture of feature-level + decision-level is constructed, breaking through the limitations of a single fusion level and effectively utilizing spatial diversity gain. At the feature level, DCA is used to fuse features from multiple sensors to obtain strongly representative fingerprint features; at the decision level, ensemble learning is combined to perform secondary verification of the classification results. Even if some sensors are in deep fading or subjected to strong interference, the system can still achieve accurate identification through complementary information from other links, improving the robustness of the system. Attached Figure Description

[0031] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0033] Figure 2 This is a flowchart of the multi-source feature fusion radiation source identification sub-process in this invention.

[0034] Figure 3 This is a diagram showing the features extracted by sensor 1 before fusion in an embodiment of the present invention.

[0035] Figure 4 This is a diagram showing the features extracted by sensor 2 before fusion in an embodiment of the present invention.

[0036] Figure 5 This is a diagram illustrating the fused features in an embodiment of the present invention.

[0037] Figure 6 This invention presents a comparison of the identification accuracy of distributed monitoring and single-station monitoring for radiation source samples of seven categories.

[0038] Figure 7This invention presents a comparison of the identification accuracy of distributed monitoring and single-station monitoring for 10 types of radiation sources.

[0039] Figure 8 This invention presents a comparison of the identification accuracy of distributed monitoring and single-station monitoring for 15 types of radiation sources. Detailed Implementation

[0040] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0041] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0042] This invention proposes a method and system for identifying individual radiation sources through multi-source feature fusion, combining the advantages of deep learning in feature extraction with the advantages of discriminant correlation analysis in feature fusion. The embodiments of this application are described in detail below with reference to the accompanying drawings.

[0043] like Figure 1 and Figure 2 As shown in the figure, this embodiment of the invention provides a method for individual identification of radiation sources through multi-source feature fusion, including the following steps: S1. Simultaneously collect radio frequency signals emitted by the target radiation source by deploying multiple spectrum monitoring sensor nodes within the monitoring area; S2. Input the signals collected by each sensor node into a deep neural network, and extract the deep feature vector containing the radio frequency fingerprint of the radiation source through multi-layer convolution and pooling operations. S3. A discriminant correlation analysis algorithm is used to fuse the deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. Specifically, the operation involves: calculating the covariance matrix and inter-class divergence matrix between feature groups of different sensors; finding an optimal projection transformation matrix such that the transformed features maximize the correlation between different sensors while maximizing the differences between different radiation source categories (i.e., maximizing the ratio of inter-class divergence to intra-class divergence); and concatenating or summing the transformed features to obtain a fused feature vector containing complementary information from multiple perspectives.

[0044] S4. Construct an ensemble classifier model that includes multiple base classifiers. Input the fused feature vector into multiple base classifiers, use a weighted voting mechanism for joint decision-making, and output the category label of the target radiation source.

[0045] By utilizing multiple distributed sensor nodes to acquire spatial diversity gain, even if some nodes experience signal quality degradation due to multipath fading, shadowing effects, or non-line-of-sight transmission, other nodes can still capture effective radiation source fingerprint information, fundamentally overcoming the problem of low identification accuracy of single-station monitoring in complex electromagnetic environments. Sensors at different geographical locations observe the same radiation source from different angles, and the acquired radio frequency signals contain complementary fingerprint features, providing a rich data foundation for subsequent feature fusion. Traditional methods rely on manually designed feature engineering, making it difficult to fully explore the subtle hardware fingerprint differences hidden in radio frequency signals. This invention automatically learns and extracts radiation source hardware fingerprint features such as power amplifier nonlinearity, oscillator phase noise, and I / Q modulation imbalance through multi-layer convolution and pooling operations of deep neural networks, avoiding the limitations of manual feature design. Through multi-layer nonlinear transformations, deep neural networks can extract more discriminative high-level semantic features, effectively amplifying the subtle differences between different radiation sources in the feature space, providing high-quality input for subsequent fusion identification.

[0046] Existing feature fusion methods typically focus only on the correlation between features, neglecting category information. This invention introduces category label information into discriminant correlation analysis, simultaneously optimizing two objectives during the fusion process: maximizing the correlation between sensors to extract common fingerprint information from multiple perspectives, and maximizing the feature differences between categories to enhance discriminative ability. By maximizing the correlation between features from different sensors, noise interference independently introduced by each sensor node is effectively suppressed; simultaneously, through feature concatenation or summation, complementary information from multiple perspectives is preserved, making the fused features both noise-resistant and containing complete fingerprint features.

[0047] Multiple base classifiers with different inductive biases are employed, each with varying sensitivities to data distribution. Ensemble learning integrates their respective strengths, effectively mitigating the bias and overfitting risks of individual classifiers on specific samples, thus improving recognition accuracy. Weight coefficients are preset based on the accuracy of each base classifier on the validation set; classifiers with higher accuracy receive greater weight in the decision-making process, making the joint decision more reliable and further enhancing the system's recognition performance. When some sensor nodes are experiencing deep fading or strong interference, compensation is provided through complementary information from multiple sensors during the feature fusion stage, and secondary verification is performed through ensemble learning during the decision-making stage. This results in a significantly more robust overall system compared to methods using a single fusion layer.

[0048] The embodiments of the present invention are described in detail below.

[0049] An RF fingerprint feature model was built using SystemVue software, and a distributed simulation environment was established. Within the simulation space, one transmitter to be identified and three receiver nodes (sensors) were configured. , , To simulate the impact of actual geographical distribution on signal fading, the following settings were implemented: , , The distances from the transmitter were 1000 meters, 2000 meters, and 3000 meters. Non-ideal characteristics such as I / Q modulation imbalance, oscillator phase noise, and amplifier nonlinearity were introduced at the transmitter end. Ten non-ideal parameters of the simulated radiation source are listed in Table 1. Since phase noise in SystemVue software is obtained by setting the fading decibel values ​​at different distances relative to the carrier frequency and fitting the data using interpolation, the "Phase Noise" column in the table lists the relative decibel values ​​at 100Hz, 1kHz, and 100kHz from the center carrier frequency. Three receivers independently acquired the I / Q signals and divided the data into equal-length segments to construct a... Multi-source fingerprint dataset of individuals similar to radiation sources.

[0050] Table 1 Examples of parameter settings for non-ideal radiation sources

[0051] The specific implementation process includes the following main steps: Step 1: Simultaneously collect radio frequency signals emitted by the target radiation source and construct a dataset by deploying multiple spectrum monitoring sensor nodes in the monitoring area.

[0052] Step 2, Place the sensor , , The collected signals are input into a deep neural network, and the output of the penultimate layer is extracted as the original feature vector. It should be noted that the deep neural network training process involves: First, a deep neural network model is constructed, which includes multiple convolutional layers, pooling layers, fully connected layers, and an output layer. Convolutional layers are used to extract local features of the signal, pooling layers are used for dimensionality reduction and to enhance translation invariance, fully connected layers are used to learn the combination relationships between features, and the output layer uses the Softmax function to output the class probability distribution.

[0053] Next, a large amount of sample data of known radiation source categories is collected to construct a training dataset. The training samples are input into a deep neural network, and the network output is calculated through forward propagation. The error between the predicted result and the true label is calculated using the cross-entropy loss function, the gradient of the parameters of each layer is calculated using the backpropagation algorithm, and the network parameters are updated using the Adam optimizer. The training is repeated iteratively until the model converges.

[0054] After training, the trained deep neural network is used as a feature extractor. In practical applications, the real-time signal collected by the sensor is input into the network, and the output of the penultimate layer is extracted as a deep feature vector. This feature vector contains the radio frequency fingerprint characteristics of the individual radiation source, including hardware fingerprint information such as power amplifier nonlinearity, oscillator phase noise, and I / Q modulation imbalance.

[0055] Step 3: Use discriminant correlation analysis algorithm to fuse the deep feature vectors from different sensor nodes to obtain fused feature vectors; the fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. To fully exploit the complementarity between data collected from different sensors and enhance the discriminative power of extracted features at the feature layer, this invention employs discriminant correlation analysis (DCI) for multi-source feature fusion. The DCA algorithm utilizes class label information to reduce the correlation between different categories within a feature set while enhancing the correlation between similar features across different feature sets. Multi-feature set fusion can be generalized through pairwise fusion; therefore, only dual-feature set fusion is illustrated here. For two or more sensor nodes, pairwise fusion can be performed sequentially.

[0056] Step 3-1: Construction and decorrelation of the inter-class scatter matrix. Let the eigenvectors of the two sensor nodes be respectively... and ,in For feature dimension, Let be the total number of samples. First, calculate the mean of each feature set at the category level. Let the th feature set be... Radiation-like sources in the feature set The sample mean in is The global mean is .calculate The inter-class scatter matrix:

[0057] in, The inter-class scatter matrix, This represents the total number of radiation source categories. For the first Number of samples per class For the first Class sample mean, This is the global mean. Because... It is usually not full rank, with the largest rank being 1. Redundant dimensions that do not contain discriminative information can be removed by eigenvalue decomposition. Diagonalization can then be performed to find the transformation matrix. , so that:

[0058]

[0059] Apply this method to another feature set Perform the same process and solve for its transformation matrix. ,right Normalization yields the transformed feature set. .

[0060] Step 3-2: Maximizing and Aligning Inter-set Feature Correlation. This involves maximizing and aligning the inter-set covariance matrix of the transformed feature set. Perform singular value decomposition to diagonalize:

[0061] At this point, let the quadratic transformation matrix be... , Then we have:

[0062] For the inter-set covariance matrix Normalize and obtain the final feature transformation set. , :

[0063]

[0064] in , This is the final transformation matrix.

[0065] Step 3-3: Fuse and output the transformed features. For the transformed feature vectors, two fusion methods can be used: concatenated fusion and additive fusion. Concatenated fusion increases the dimension of the fused feature vector, while additive fusion keeps the dimension unchanged.

[0066] The specific calculation method for cascade fusion is as follows:

[0067] The specific calculation method for addition and fusion is as follows:

[0068] Step 4: Construct an ensemble classifier model that includes multiple base classifiers. Input the fused feature vector into multiple base classifiers, use a weighted voting mechanism to make joint decisions, and output the category label of the target radiation source.

[0069] After obtaining highly discriminative fusion feature vectors through discriminant correlation analysis, this invention does not rely on a single classification model. Instead, it constructs a joint decision architecture based on ensemble learning. This architecture aims to effectively reduce the risk of overfitting by integrating the complementary effects of multiple heterogeneous base classifiers. Furthermore, it mitigates the impact of outliers and data perturbations on base classifiers through multi-faceted evaluation, thereby improving overall recognition accuracy and robustness. Specifically, this includes: Step 4-1: Base Classifier Configuration. To fully capture the nonlinear distribution characteristics of fused features in high-dimensional space, this embodiment selects three algorithms with different inductive biases as base classifiers: K-Nearest Neighbors (KNN), Decision Tree (DT), and Logistic Regression (LR).

[0070] Deployment process of ensemble learning models: During the training phase, the fused feature vectors output from step 3 and their corresponding class labels are used to construct a training set. Three base classifiers—K-Nearest Neighbors, Decision Tree, and Logistic Regression—are trained separately. The K-Nearest Neighbors classifier uses Euclidean distance as the distance metric, and the K value is determined through cross-validation. The Decision Tree classifier is constructed using the CART algorithm, and splitting features are selected using the Gini coefficient. The Logistic Regression classifier uses L2 regularization to prevent overfitting and optimizes parameters using gradient descent.

[0071] The accuracy of each base classifier is evaluated on the validation set, and weight coefficients are preset based on the accuracy, with higher accuracy resulting in greater weights.

[0072] In the application phase, the three trained base classifiers are loaded into the memory of the signal processing unit (the classification and recognition module in this embodiment). When the fused feature vector output from step 3 is input, the three base classifiers infer in parallel and output class probability distribution vectors respectively. Finally, the votes of each classifier are weighted and summed according to preset weight coefficients, and the class with the highest weighted score is selected as the final recognition result.

[0073] Step 4-2, Decision Generation. The fused feature vectors are simultaneously input into the base classifiers described above. Each classifier outputs its predicted class label and confidence score. To preserve uncertainty during the prediction process, each learner does not directly output hard class labels, but instead outputs a class score vector that follows a posterior probability distribution. For the... Individual learners Its output can be represented as a probability distribution vector:

[0074] in, Indicates that under known fusion features And the model is Under the conditions, the target radiation source belongs to the first Confidence level of radiation-like sources.

[0075] Step 4-3: Voting Decision. A weighted majority voting method is used to determine the final result. Assume each classifier has a certain impact on the categories. The vote was Final recognition result for:

[0076] in These are the preset weight coefficients based on the accuracy of each classifier on the validation set.

[0077] This invention extracts sensor data during the training process separately. , The contour coefficients of the feature vectors extracted by the neural network model and the fused feature vectors were experimentally observed against the visualized feature maps. The feature contour coefficients before and after fusion are shown in Table 2. Figure 3 , Figure 4 and Figure 5 As shown, the intra-class aggregation degree and inter-class differences of fused features are significantly higher than those of single features. (Sensor) In the feature space, categories "3" and "8", and "5" and "7" are too close together, indicating that the sensor... In the feature space, there is even some overlap between category "3" and category "8", but the situation is greatly improved for fused features.

[0078] Table 2 Comparison of feature contour coefficients before and after fusion

[0079] This invention targets six radar experimental signal test sets: Barker Code, CW, Frank Code, LFM, Polytime Code, and Pulse signals. Under conditions of 7, 10, and 15 radiation source sample categories, and 3, 5, 10, 15, and 20 sample signals respectively, based on single-station and distributed environments, and single-classifier and ensemble-classifier environments, a 3-fold cross-validation method is used to conduct 100 Monte Carlo simulations. The recognition accuracy under different sample conditions and experimental methods is then compared. Figure 6 , Figure 7 and Figure 8 As shown, both multi-sensor feature fusion and multi-classifier joint decision-making can effectively improve the performance of radiation source identification. Moreover, as the number of radiation source categories to be identified increases, the performance improvement of distributed monitoring becomes increasingly significant when single-station monitoring is inadequate. When there are 10 and 15 radiation source categories, distributed monitoring shows an improvement of approximately 8% and 12% compared to single-station monitoring, which is sufficient to demonstrate the effectiveness of the proposed method.

[0080] This invention proposes a method that, at the feature level, fuses features extracted from multiple sensors via a neural network, effectively achieving feature aggregation within classes and differentiation between classes; at the decision level, it fuses decisions made by multiple classifiers. This improves the final recognition accuracy. Experimental examples are used to quantitatively evaluate and verify the effectiveness of the proposed method using contour coefficients, visualized feature maps, and recognition accuracy.

[0081] Traditional feature fusion methods only focus on the correlation between features, while this application introduces category information simultaneously during the fusion process through the DCA algorithm, resulting in an ideal distribution of fused features in the feature space: intra-class aggregation and inter-class dispersion. Experiments show that the silhouette coefficient of the fused features is improved by approximately 3-4 percentage points. Feature-level fusion and decision-level fusion are not simply superimposed but form a synergistic effect. Feature-level fusion provides high-quality input for the decision-level fusion, while decision-level fusion performs secondary optimization on the feature-level output. Experiments show that when the number of radiation source categories is 15, the recognition accuracy of this invention is improved by approximately 12% compared to single-station monitoring. When some sensor nodes are in deep fading or subjected to strong interference, the system can still achieve accurate identification through complementary information from other links, effectively overcoming the problem of single-station monitoring being greatly affected by channel conditions.

[0082] This invention also provides a multi-source feature fusion radiation source individual identification system, comprising: Multiple distributed spectrum monitoring sensor nodes are deployed within the monitoring area to synchronously acquire radio frequency signals emitted by the target radiation source; The feature extraction module is equipped with a deep neural network, which is used to input the signals collected by each sensor node into the deep neural network and extract deep feature vectors containing the radio frequency fingerprint of the radiation source through multi-layer convolution and pooling operations. The feature fusion module employs a discriminant correlation analysis algorithm to fuse deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. The classification and recognition module is equipped with an ensemble classifier model, which includes multiple base classifiers. The ensemble classifier model is used to input the fused feature vector into the multiple base classifiers, and to make joint decisions using a weighted voting mechanism, and output the category label of the target radiation source. The output module is used to display or report the category labels.

[0083] In some embodiments, the feature fusion module employs a discriminant correlation analysis algorithm to fuse deep feature vectors from different sensor nodes, specifically for: The inter-class scatter matrix is ​​calculated for the feature set of each sensor node, and the transformation matrix is ​​obtained through eigenvalue decomposition. The feature set is then subjected to decorrelation transformation. Calculate the inter-set covariance matrix for multiple feature sets after decorrelation transformation, obtain the quadratic transformation matrix through singular value decomposition, and perform correlation maximization transformation on the feature sets. The features of the correlation-maximizing transformation are fused to generate a fused feature vector.

[0084] The inter-class scatter matrix is ​​calculated as follows:

[0085] The correlation maximization transformation includes: For the two feature sets after decorrelation transformation and Calculate its inter-set covariance matrix. ;right Perform singular value decomposition Let the quadratic transformation matrix be , Obtain the final transformed features , .

[0086] The fusion of features that maximize correlation transformation includes: Using serial fusion method Alternatively, an additive fusion method may be adopted. .

[0087] In some embodiments, the ensemble classifier model in the classification and recognition module includes three base classifiers: K-nearest neighbors, decision trees, and logistic regression; the classification and recognition module is specifically used for: The fused feature vector is simultaneously input into each base classifier, and each base classifier outputs a class probability distribution vector. A weighted majority voting method is used to jointly decide the outputs of each base classifier to determine the final recognition result.

[0088] Probability distribution vector:

[0089] in, Indicates that under known fusion features And the model is Under the conditions, the target radiation source belongs to the first Confidence level of radiation-like sources.

[0090] The weighted majority voting method is as follows: Weight coefficients are preset based on the accuracy of each base classifier on the validation set. The votes of each base classifier for each category are weighted and summed, and the category with the highest weighted score is selected as the final recognition result. The formula is:

[0091] in, For the final identification result, For indicator functions, For the first A base classifier for each category The vote These are the preset weight coefficients based on the accuracy of each base classifier on the validation set.

[0092] Take the category corresponding to the maximum value. Among the C categories, select the category c that makes the subsequent expression have the largest value.

[0093] The system constructs a multi-level fusion architecture at the feature level and the decision level: at the feature level, features from multiple sensors are fused through discriminant correlation analysis; at the decision level, classification results are jointly decided through ensemble learning.

[0094] The output module includes a display device and a communication interface, used to visualize the category labels or report them to the spectrum monitoring center.

[0095] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or other media capable of storing program code. It includes several instructions to cause a computer terminal (which may be a personal computer, server, or a second terminal, network terminal, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0096] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0097] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0098] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0099] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for individual identification of radiation sources through multi-source feature fusion, characterized in that, Includes the following steps: The radio frequency signals emitted by the target radiation source are collected synchronously by deploying multiple spectrum monitoring sensor nodes within the monitoring area. The signals collected by each sensor node are input into a deep neural network, and deep feature vectors containing the radio frequency fingerprint of the radiation source are extracted through multi-layer convolution and pooling operations. A discriminant correlation analysis algorithm is used to fuse deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. An ensemble classifier model comprising multiple base classifiers is constructed. The fused feature vector is input into multiple base classifiers, and a weighted voting mechanism is used for joint decision-making to output the category label of the target radiation source.

2. The method for individual identification of multi-source feature fusion radiation sources according to claim 1, characterized in that, Discriminant correlation analysis algorithm is used to fuse deep feature vectors from different sensor nodes, specifically including: The inter-class scatter matrix is ​​calculated for the feature set of each sensor node, and the transformation matrix is ​​obtained through eigenvalue decomposition. The feature set is then subjected to decorrelation transformation. Calculate the inter-set covariance matrix for multiple feature sets after decorrelation transformation, obtain the quadratic transformation matrix through singular value decomposition, and perform correlation maximization transformation on the feature sets. The features of the correlation-maximizing transformation are fused to generate a fused feature vector.

3. The method for individual identification of multi-source feature fusion radiation sources according to claim 2, characterized in that, The inter-class scatter matrix is ​​calculated as follows: in, The inter-class scatter matrix, This represents the total number of radiation source categories. For the first Number of samples per class For the first Class sample mean, This is the global mean.

4. The method for individual identification of multi-source feature fusion radiation sources according to claim 2, characterized in that, The correlation maximization transformation includes: For the two feature sets after decorrelation transformation and Calculate its inter-set covariance matrix. ;right Perform singular value decomposition Let the quadratic transformation matrix be , Obtain the final transformed features , .

5. The method for individual identification of multi-source feature fusion radiation sources according to claim 4, characterized in that, The fusion of features that maximize correlation transformation includes: Using serial fusion method Alternatively, an additive fusion method may be adopted. .

6. The method for individual identification of multi-source feature fusion radiation sources according to claim 1, characterized in that, The ensemble classifier model includes three base classifiers: K-nearest neighbors, decision trees, and logistic regression. The fused feature vector is input into the ensemble classifier model, which outputs the class label of the target radiation source, specifically including: The fused feature vector is simultaneously input into each base classifier, and each base classifier outputs a class probability distribution vector. A weighted majority voting method is used to jointly decide the outputs of each base classifier to determine the final recognition result.

7. The method for individual identification of multi-source feature fusion radiation sources according to claim 6, characterized in that, Probability distribution vector: in, Indicates that the fusion features are known And the model is Under the conditions, the target radiation source belongs to the first Confidence level of radiation-like sources.

8. The method for individual identification of multi-source feature fusion radiation sources according to claim 6, characterized in that, The weighted majority voting method is as follows: Weight coefficients are preset based on the accuracy of each base classifier on the validation set. The votes of each base classifier for each category are weighted and summed, and the category with the highest weighted score is selected as the final recognition result. The formula is: in, For the final identification result, For indicator functions, For the first A base classifier for each category The vote These are the preset weight coefficients based on the accuracy of each base classifier on the validation set.

9. The method for individual identification of multi-source feature fusion radiation sources according to claim 1, characterized in that, The steps of constructing an ensemble classifier model that includes multiple base classifiers, inputting the fused feature vector into multiple base classifiers, using a weighted voting mechanism for joint decision-making, and outputting the class label of the target radiation source include: The category labels can be visualized via a display device or reported to the spectrum regulatory center via a communication interface.

10. A multi-source feature fusion radiation source individual identification system, characterized in that, include: Multiple distributed spectrum monitoring sensor nodes are deployed within the monitoring area to synchronously acquire radio frequency signals emitted by the target radiation source; The feature extraction module is equipped with a deep neural network, which is used to input the signals collected by each sensor node into the deep neural network and extract deep feature vectors containing the radio frequency fingerprint of the radiation source through multi-layer convolution and pooling operations. The feature fusion module employs a discriminant correlation analysis algorithm to fuse deep feature vectors from different sensor nodes to obtain a fused feature vector. The fusion process maximizes the feature correlation between different sensors while maximizing the feature differences between different radiation source categories. The classification and recognition module is equipped with an ensemble classifier model, which includes multiple base classifiers. The ensemble classifier model is used to input the fused feature vector into the multiple base classifiers, and to make joint decisions using a weighted voting mechanism, and output the category label of the target radiation source. The output module is used to display or report the category labels.

Images