A method for individual identification of communication radiation sources based on multi-scale residual prototype learning networks

Abstract

This invention relates to a method for identifying individual communication radiation sources based on a multi-scale residual prototype learning network. The method comprises a signal preprocessing module, a network training and feature extraction module, a storage module, and an identification module. The signal preprocessing module is connected to the network training and feature extraction module, which is connected to the storage module, and the storage module is connected to the identification module. This invention extracts features from individual communication radiation source signals using a multi-scale residual prototype learning network and then uses a joint decision-making method to identify the attributes of the individual radiation source signals. In this method, MSRPLNet lays the foundation for identifying the features of individual communication radiation source signals by enhancing the inter-class separability and intra-class compactness of the learned features.

Description

Technical Field

[0001] This invention belongs to the field of space situational awareness and electronic countermeasures technology, specifically involving a method for individual identification of communication radiation sources based on a multi-scale residual prototype learning network. Background Technology

[0002] With the increasing number of various types of communication radiation sources (such as WiFi, Zigbee, etc.), the management of limited spectrum resources is becoming increasingly difficult. Malicious spectrum access severely impacts the normal use and experience of spectrum resources for primary users. Furthermore, malicious transmitters may be used to deceive users and affect the secure transmission of wireless information. Therefore, we need to take effective measures to identify trustworthy, harmful, and new communication radiation sources. However, effectively sensing and identifying a massive number of individual radiation sources remains a challenging problem.

[0003] Existing methods for identifying individual radiation sources are primarily experience-driven and data-driven. Experience-driven methods first rely on manual extraction of intuitive and reliable expert features from the radiation source signal. These features are then matched against templates in a radiation source feature library, and the attribute of the template with the highest similarity is the identification result. Expert features are mainly divided into modulation domain features and waveform domain features. Modulation domain features mainly include I / Q amplitude and phase imbalances in the received signal caused by small-scale hardware defects, carrier frequency offset, and modulation offset. Common waveform domain features include compressed bispectrum, fractal box dimension, multi-scale approximate entropy, Hilbert-Huang transform, empirical mode decomposition, and variational mode decomposition. However, experience-driven methods rely on prior knowledge of the signal and existing signal processing tools. Furthermore, the extracted expert features are easily affected by noise, leading to significant fluctuations in identification accuracy. In practical applications, experience-driven methods are often constrained by the real-world environment and rarely achieve satisfactory results.

[0004] With the development of deep learning theory and the parallel and fast computing support provided by GPUs, data-driven identification of individual radiation sources has become a research hotspot. Intelligent identification technologies based on deep learning algorithms can automatically extract radiation source signal features through network structure design and acquire adaptive models through training on massive amounts of data. However, most deep learning work in this field essentially only improves the inter-class separability of radiation source signal features. They often neglect to enhance the intra-class compactness of radiation source signal features. Furthermore, the openness of real-world environments may prevent existing closed-set identification methods from detecting and rejecting new, unknown radiation sources, significantly increasing the difficulty of identifying communication radiation sources. Summary of the Invention

[0005] The purpose of this invention is to solve the above-mentioned problems and provide a method for individual identification of communication radiation sources based on a multi-scale residual prototype learning network, which can accurately classify known radiation sources while also detecting and rejecting unknown radiation sources.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for individual identification of communication radiation sources based on a multi-scale residual prototype learning network. The method comprises a signal preprocessing module, a network training and feature extraction module, a storage module, and an identification module. The signal preprocessing module is connected to the network training and feature extraction module, the network training and feature extraction module is connected to the storage module, and the storage module is connected to the identification module.

[0007] The signal preprocessing module preprocesses the intermediate frequency signal of the acquired original communication radiation source, including signal detection and subband signal separation, slicing, denoising, using Hilbert transform to obtain I and Q signals, power normalization / standard normalization, setting labels, and constructing a three-dimensional sample dataset, which includes a training set, a validation set, and a test set.

[0008] The network training and feature extraction module is trained using a multi-scale residual prototype learning network (MSRPLNet) and automatically extracts individual signal features of communication radiation sources. The multi-scale residual prototype learning network consists of an input layer, a convolutional layer, a batch normalization layer, an average pooling layer, a ReLU activation layer, a packet loss layer, a multi-scale residual block, a flattening layer, and a fully connected layer, and is trained using a prototype learning-based strategy.

[0009] The storage module is used to store the trained MSRPLNet parameter file on the computer for later use.

[0010] The identification module uses a joint decision-making method to authenticate the new data samples. The identification results are of two kinds: known communication radiation sources and their corresponding categories, new communication radiation sources, or unknown communication radiation sources.

[0011] Furthermore: The input of the MSRPLNet is a slice of intermediate frequency I / Q signal after standard normalization. The input layer first uses a 1×9 convolutional kernel to extract features from the original data; the batch normalization layer is used to standardize the data of each channel to stabilize the data distribution of each layer of the network and speed up the learning speed of the network; the average pooling layer is set to a size of 2×2, which can merge the features in each channel into one channel, and the number becomes one-quarter of the original; the activation layer (ReLU) is used to perform non-linear processing on the output of the upper layer neurons and pass it to the lower layer neurons; the dropout layer drops some neurons to prevent the model from overfitting, and then inputs the extracted shallow features into multiple consecutive multi-scale residual blocks.

[0012] Furthermore: In the multi-scale residual block, the first feature is extracted by two 1×7 convolution kernels, the second feature is extracted by two 1×5 convolution kernels, the third feature is extracted by two 1×3 convolution kernels, and the fourth feature is the original feature, extracted by two 1×1 convolution kernels.

[0013] Furthermore, to further enhance inter-class separability and intra-class compactness of features, MSRPLNet introduces a prototype learning strategy as a model training method. During training, MSRPLNet maintains and learns a prototype for each class, denoted as m. ij , where i∈{1,2,...,M} represents the category index, M represents the number of categories, j={1,2,...,K} represents the index of the prototype in each class, and K represents the number of prototypes in each class;

[0014] The joint loss based on prototype learning is obtained by weighting the distance cross-entropy loss (DCE Loss) and the prototype loss (PL Loss). Let f(·) be a carefully designed feature extractor. Since distance can be used to measure similarity, sample x belongs to prototype m. ij The probability is represented by distance, i.e.

[0015]

[0016] Similar to Softmax Loss, DCE Loss also uses the Softmax function to calculate the output probability, i.e.

[0017] p(x∈m i,j |x,f)=Softmax[-d(f(x),m ij ) / γ] (2)

[0018] In the formula Representative sample feature f(x) and prototype m ij The Euclidean distance between them, γ is a hyperparameter used to control distance stiffness, therefore, the DCE Loss of sample x belonging to the y-th class can be expressed as:

[0019]

[0020] Considering that directly optimizing the DCE loss may lead to model overfitting, PL Loss is introduced as a regularization term to improve the model's generalization ability. PL Loss is defined as follows:

[0021]

[0022] In the formula m yj The Joint Loss is obtained by weighting the DCE Loss and PL Loss to represent the nearest prototype of category y corresponding to f(x).

[0023] l Joint ((x,y); f; M)=l DCE ((x,y);f;M)+λ·l PL ((x,y);f;M) (5)

[0024] In the formula, λ is a hyperparameter used to balance DCE Loss and PL Loss.

[0025] Furthermore: In the joint decision-making method, it is assumed that each category has only one prototype. Let f(x) be the feature extracted from sample x using MSRPLNet, and m be the prototype with the smallest Euclidean distance to f(x). i m i The central feature of the corresponding i-th class is c i c i It is obtained by averaging the features of correctly identified samples in the validation set. Given thresholds δ1 and δ2, when f(x) is respectively related to m i and c i When the distance satisfies the following relationship

[0026]

[0027] Then x is considered to belong to the prototype representation m i The corresponding i-th class is determined; otherwise, it is determined to be an unknown radiation source sample.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] The proposed method for identifying individual communication radiation sources uses a multi-scale residual prototype learning network to extract features from individual communication radiation source signals, and then employs a joint decision-making method to identify the attributes of the individual communication radiation source. Within this framework, MSRPLNet lays the foundation for identifying unknown radiation source targets by enhancing the inter-class separability and intra-class compactness of the learned features. MSRPLNet completely abandons the Softmax layer and transforms probability-based discrimination into distance- or similarity-based discrimination, solving the closure problem inherent in the Softmax layer. Prototypes in MSRPLNet can be learned together with network parameters without requiring predefined rules for updates, significantly enhancing the learnability of prototypes and their ability to represent class features. Furthermore, unlike traditional neural networks (such as CNNs) that learn features, MSRPLNet does not partition the entire feature space but projects sample features onto the vicinity of the prototype. Therefore, MSRPLNet, based on prototype learning, has better detection and rejection capabilities for unknown samples. Most existing open-set recognition methods are based on probability thresholds and generative adversarial networks (GANs). Probability thresholds are generally difficult to determine and often rely on experience; the same threshold is hard to adapt to different data. The training process for GANs is too complex and struggles to meet the real-time requirements of related fields. In contrast, a joint spatial target discrimination method based on prototype distance and center distance can automatically set the threshold, which is directly derived from the data characteristics without manual setting. Furthermore, the threshold can adaptively adjust when the data distribution of individual radiation source signals changes, further improving the generalization ability of the joint recognition method. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only for more clearly illustrating the technical solutions in the embodiments of the present invention or the prior art. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is an overall framework diagram of the present invention;

[0032] Figure 2 This is a schematic diagram of the MSRPLNet multi-scale residual prototype learning network of the present invention;

[0033] Figure 3 This is a schematic diagram of the multi-scale residual fastness of the present invention;

[0034] Figure 4 This is a schematic diagram of the joint decision-making method of the present invention; Detailed Implementation

[0035] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be further described below with reference to specific embodiments. However, the embodiments are only for illustration and are not intended to limit the present invention.

[0036] like Figure 1 The method for individual identification of communication radiation sources based on multi-scale residual prototype learning networks is shown. The method consists of a signal preprocessing module, a network training and feature extraction module, a storage module, and an identification module. The signal preprocessing module is connected to the network training and feature extraction module, the network training and feature extraction module is connected to the storage module, and the storage module is connected to the identification module.

[0037] The signal preprocessing module preprocesses the acquired intermediate frequency signal from the original communication radiation source, including signal detection and subband signal separation. The signal detection and subband signal separation module is connected to the signal receiving module. The transmitted data is sliced, denoised, and the I and Q signals are obtained using Hilbert transform. The power is normalized / standardized, labels are set, and a three-dimensional sample dataset is constructed. The three-dimensional sample dataset includes a training set, a validation set, and a test set.

[0038] The network training and feature extraction module trains the Multi-Scale Residual Prototype Learning Network (MSRPLNet) of this invention and automatically extracts the features of space communication radiation source signals. For example... Figure 2 As shown, the multi-scale residual prototype learning network MSRPLNet mainly consists of an input layer, convolutional layers, batch normalization layers, average pooling layers, ReLU activation layers, packet loss layers, multi-scale residual blocks, flattening layers, and fully connected layers, and is trained using a prototype learning-based strategy. Convolutional kernels of different scales are used to learn inter-class separation features. The residual network architecture serves two purposes: mitigating gradient vanishing and model degradation, and fusing features from different scales. Optimizing the joint loss based on prototype learning not only enhances the inter-class separability of features but also further enhances the intra-class compactness of features.

[0039] The input to MSRPLNet is a slice of mid-frequency I / Q signal after standard normalization. It first uses a 1×2 convolutional kernel to extract features from the raw data. Larger kernels have a larger receptive field, allowing for better observation of the information contained in the data and extraction of global features. The batch normalization layer standardizes the data for each channel to stabilize the data distribution in each layer of the network and accelerate the learning speed. Considering that the extracted features would become too large after using a flattening layer, average pooling is used on the feature maps in the early stages to reduce computation and memory requirements. The pooling size is set to 2×2, meaning that the features in each channel will be merged into one channel, reducing the number of features to one-quarter of the original. The activation layer (ReLU) performs non-linear processing on the output of the upper-layer neurons and passes it to the lower-layer neurons. The dropout layer prevents overfitting by discarding some neurons and then inputs the extracted shallow features into multiple consecutive multi-scale residual blocks.

[0040] like Figure 3 As shown, the multi-scale residual blocks share the same structure, inspired by the residual units in residual neural networks. In a multi-scale residual block, the first feature path is extracted by two 1×7 convolutional kernels; the second feature path is extracted by two 1×5 convolutional kernels; the third feature path is extracted by two 1×3 convolutional kernels; and the fourth feature path is the original feature, which can be considered as being extracted by two 1×1 convolutional kernels. Different sized convolutional kernels possess different receptive fields, allowing them to capture more differentiated hierarchical features. The four features at different scales are then processed by an Add layer to obtain the output features. On one hand, the Add layer inherits the advantages of residual units and can be used to alleviate gradient vanishing and model degradation problems. On the other hand, the Add layer fuses the four features, resulting in features with greater inter-class separability.

[0041] After several consecutive multi-scale residual blocks, MSRPLNet can extract deep features from signals of different spatial radiation sources. The number of multi-scale residual blocks used depends on the network's learning progress. Finally, a flattening layer and three fully connected layers with different numbers of neurons are used to integrate the extracted abstract features to obtain the final output features. During the training of MSRPLNet, the Adam optimizer is used to optimize the prototype-based joint loss to update the network parameters and prototype representation.

[0042] To further enhance the inter-class separability and intra-class compactness of features, MSRPLNet introduces a prototype learning strategy as a model training method. During training, MSRPLNet maintains and learns several prototypes for each class. Let m be the prototypes. ijWhere i∈{1,2,...,M} represents the class index, M represents the number of classes, j={1,2,...,K} represents the index of the prototype in each class, and K represents the number of prototypes in each class. The prototype representation can be seen as an abstract representation of the features and can be learned along with the features. The Joint Loss based on prototype learning is obtained by weighting the Distance Cross Entropy Loss (DCE Loss) and the Prototype Loss (PL Loss). The DCE Loss is very similar to the Classification Cross Entropy Loss (Softmax Loss), the difference being that the Softmax Loss optimizes the similarity between the features and the true labels, while the DCE Loss optimizes the similarity between the features and the true prototypes.

[0043] Let f(·) be a carefully designed feature extractor. Since distance can be used to measure similarity, sample x belongs to prototype m. ij The probability can be represented by distance, i.e.

[0044]

[0045] Similar to Softmax Loss, DCE Loss also uses the Softmax function to calculate the output probability, i.e.

[0046] p(x∈m i,j |x,f)=Softmax[-d(f(x),m ij ) / γ] (2)

[0047] In the formula Representative sample feature f(x) and prototype m ij The Euclidean distance between them, γ is a hyperparameter used to control distance stiffness. Therefore, the DCE Loss of sample x belonging to class y can be expressed as:

[0048]

[0049] Considering that directly optimizing the DCE loss may lead to model overfitting, PL Loss is introduced as a regularization term to improve the model's generalization ability. PL Loss is defined as:

[0050]

[0051] In the formula m yj This represents the nearest prototype of category y corresponding to f(x);

[0052] By weighting the DCE Loss and PL Loss, we can further obtain the Joint Loss:

[0053] l Joint ((x,y); f; M)=l DCE ((x,y);f;M)+λ·l PL ((x,y);f;M) (5)

[0054] In the formula, λ is a hyperparameter used to balance DCE Loss and PL Loss.

[0055] Both DCE Loss and PL Loss are differentiable for both prototypes and network parameters, meaning that during training, prototypes and network parameters can be jointly learned by optimizing Joint Loss. DCE Loss inherits the characteristics of Softmax Loss and can be used to learn inter-class separation features; PL Loss enhances intra-class compactness by penalizing the distance between features and their corresponding prototypes. Furthermore, since the prototypes are learnable, the optimal prototype distribution can be found in the high-dimensional feature space after MSRPLNet training, which plays a decisive role in improving the recognition performance of individual communication radiation sources.

[0056] The storage module is used to store the trained MSRPLNet parameter file on the computer for later use. When recognizing new data samples, the parameter file can be directly called for testing.

[0057] The identification module uses a joint decision-making method to authenticate the new data samples. The identification results are of two types: 1. Known communication radiation sources and their corresponding categories; 2. New communication radiation sources / unknown communication radiation sources.

[0058] like Figure 4 The diagram illustrates the joint decision-making method. Assuming each class has only one prototype, let f(x) be the feature extracted from sample x using MSRPLNet, and let m be the prototype with the smallest Euclidean distance to f(x). i m i The central feature of the corresponding i-th class is c i c i This can be obtained by averaging the features of correctly identified samples in the validation set. Given thresholds δ1 and δ2, when f(x) is related to m... i and c i When the distance satisfies the following relationship

[0059]

[0060] x can be considered to belong to the prototype representation m iThe corresponding i-th class. Conversely, the sample is classified as an unknown radiation source sample.

[0061] Regarding how to determine the thresholds δ1 and δ2, the following method is provided for setting the thresholds:

[0062] (1) Once the network validation accuracy reaches the expected level, the network training can be considered complete; for any correctly identified sample x in the validation set, output its corresponding feature f(x) and the nearest prototype m. i And classify it into the i-th category;

[0063] (2) After classification, calculate the central feature c of the i-th type of radiation source sample. i ;

[0064] (3) Find the sample that is furthest from its prototype among the correctly identified samples of each class and calculate the corresponding prototype distance.

[0065] (4) Find the sample farthest from the center among the correctly identified samples of each class and calculate the corresponding center distance.

[0066] (5) Order Thresholds δ1 and δ2 can be automatically obtained from the data;

[0067] During the testing phase, we first use MSRPLNet to extract the features f(z) of the test sample z, and then calculate the distance d between f(z) and its nearest prototype. mb And the distance d between f(z) and its nearest center cb ,like

[0068] d mb ≤δ1andd cb ≤δ2 (7)

[0069] If the test sample belongs to a known space radiation source, it is classified as class b; otherwise, the test sample is classified as an unknown class.

[0070] The proposed framework for identifying space communication radiation sources uses a multi-scale residual prototype learning network to extract features from communication radiation source signals, and then employs a joint decision-making method to identify the attributes of the communication radiation sources. In this framework, MSRPLNet lays the foundation for identifying unknown space communication radiation sources by enhancing the inter-class separability and intra-class compactness of the learned features. It completely abandons the Softmax layer and transforms probability-based discrimination into distance- or similarity-based discrimination, solving the closure problem inherent in the Softmax layer. Prototypes in MSRPLNet can be learned together with network parameters without requiring predefined rules for updates, greatly enhancing the learnability of prototypes and their ability to represent class features. Furthermore, unlike traditional neural networks (such as CNNs) that learn features, MSRPLNet does not partition the entire feature space but projects sample features onto the vicinity of the prototype. Therefore, prototype-based MSRPLNet has better detection and rejection capabilities for unknown samples. Most existing open-set identification methods are based on probability thresholds and generative adversarial networks (GANs). Probability thresholds are generally difficult to determine and often rely on experience; the same threshold is difficult to adapt to different data. The training process of GANs is too complex and struggles to meet the real-time requirements of related fields. The joint discrimination method for space communication radiation sources based on prototype distance and center distance can automatically set a threshold, which is directly obtained from the characteristics of the data without manual setting. Furthermore, the threshold can adaptively adjust when the data distribution of space communication radiation source signals changes, further improving the generalization ability of the joint identification method.

[0071] All content not described in detail in this invention is prior art.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for individual identification of communication radiation sources based on multi-scale residual prototype learning networks, characterized in that: The method consists of a signal preprocessing module, a network training and feature extraction module, a storage module, and a recognition module. The signal preprocessing module is connected to the network training and feature extraction module, the network training and feature extraction module is connected to the storage module, and the storage module is connected to the recognition module. The signal preprocessing module preprocesses the intermediate frequency signal of the acquired original communication radiation source, including signal detection and subband signal separation, slicing, denoising, using Hilbert transform to obtain I and Q signals, power normalization / standard normalization, setting labels, and constructing a three-dimensional sample dataset, which includes a training set, a validation set, and a test set. The network training and feature extraction module is trained using a multi-scale residual prototype learning network (MSRPLNet) and automatically extracts individual signal features of communication radiation sources. The multi-scale residual prototype learning network consists of an input layer, a convolutional layer, a batch normalization layer, an average pooling layer, a ReLU activation layer, a packet loss layer, a multi-scale residual block, a flattening layer, and a fully connected layer, and is trained using a prototype learning-based strategy. The storage module is used to store the trained MSRPLNet parameter file on the computer for later use. The identification module uses a joint decision-making method to authenticate the new data samples. The identification results are of two types: known communication radiation sources and their corresponding categories; new communication radiation sources or unknown communication radiation sources. The input to MSRPLNet is a slice of mid-frequency I / Q signal after standard normalization. The input layer first uses a 1×9 convolutional kernel to extract features from the original data. The batch normalization layer is used to standardize the data of each channel to stabilize the data distribution of each layer of the network and speed up the learning speed of the network. The average pooling layer is set to a size of 2×2, which merges the features in each channel into one channel, and the number is reduced to one-quarter of the original. The activation layer (ReLU) is used to non-linearly process the output of the upper layer neurons and pass it to the lower layer neurons. The dropout layer prevents the model from overfitting by discarding some neurons, and then inputs the extracted shallow features into multiple consecutive multi-scale residual blocks. In the multi-scale residual block, the first feature is extracted by two 1×7 convolution kernels, the second feature is extracted by two 1×5 convolution kernels, the third feature is extracted by two 1×3 convolution kernels, and the fourth feature is the original feature, which is extracted by two 1×1 convolution kernels.

2. The method for individual identification of communication radiation sources based on multi-scale residual prototype learning networks according to claim 1, characterized in that: To further enhance the inter-class separability and intra-class compactness of learned features, MSRPLNet introduces a prototype learning strategy as a model training method. During training, MSRPLNet maintains and learns a prototype for each class, denoted as m. ij , where i∈{1,2,...,M} represents the category index, M represents the number of categories, j={1,2,...,K} represents the index of the prototype in each class, and K represents the number of prototypes in each class; The joint loss based on prototype learning is obtained by weighting the distance cross-entropy loss (DCELoss) and the prototype loss (PL Loss). Let f(·) be a carefully designed feature extractor. Since distance can be used to measure similarity, sample x belongs to prototype m. ij The probability is represented by distance, i.e. Similar to Softmax Loss, DCE Loss also uses the Softmax function to calculate the output probability, i.e. p(x∈m i,j |x,f)=Softmax[-d(f(x),m ij ) / γ] (2) In the formula Representative sample feature f(x) and prototype m ij The Euclidean distance between them, γ is a hyperparameter used to control the distance stiffness. Therefore, the DCE Loss for sample x belonging to class y can be expressed as: Considering that directly optimizing the DCE loss may lead to model overfitting, PL Loss is introduced as a regularization term to improve the model's generalization ability. PL Loss is defined as follows: In the formula m yj This represents the nearest prototype of category y corresponding to f(x). By weighting the DCE Loss and PL Loss, we obtain the Joint Loss: l Joint ((x,y);f;M)=l DCE ((x,y);f;M)+λ·l PL ((x,y);f;M) (5) In the formula, λ is a hyperparameter used to balance DCE Loss and PL Loss.

3. The method for individual identification of communication radiation sources based on multi-scale residual prototype learning networks according to claim 1, characterized in that: In the joint decision-making method, it is assumed that each category has only one prototype. Let f(x) be the feature extracted from sample x using MSRPLNet, and m be the prototype with the smallest Euclidean distance to f(x). i m i The central feature of the corresponding i-th class is c i c i It is obtained by averaging the features of correctly identified samples in the validation set. Given thresholds δ1 and δ2, when f(x) is respectively compared with m i and c i When the distance satisfies the following relationship x belongs to the prototype representation m i The corresponding i-th class is determined; otherwise, it is determined to be an unknown radiation source sample.

Images