A method and device for cross-receiver emitter individual identification based on multi-domain generalization

Abstract

The application discloses a cross-receiver radiation source individual identification method and device based on multi-domain generalization, and belongs to the technical field of wireless communication and signal processing. The method comprises the following steps: acquiring I / Q signals generated by capturing and converting wireless signals transmitted by an unknown radiation source by a receiver; after pre-processing the I / Q signals, inputting the I / Q signals into a pre-trained radiation source individual identification model, processing the I / Q signals by a feature extractor and a classifier, and outputting identity information of the unknown radiation source; wherein, the training of the identification model only adopts source domain samples. The application adopts only source domain samples to train the identification model, realizes high-precision and high-robust identification of unknown radiation source individuals, and effectively suppresses the influence of receiver differences.

Description

Technical Field

[0001] This invention relates to a method and apparatus for cross-receiver radiation source individual identification based on multi-domain generalization, belonging to the field of wireless communication and signal processing technology. Background Technology

[0002] Existing deep learning-based methods for identifying individual radiation sources typically implicitly assume that the receivers used in the training and testing phases are identical. However, with the widespread adoption of wireless communication technology, the actual deployment scenarios for radiation source identification systems are becoming increasingly complex. Signal acquisition during training and deployment may be performed by receivers of various models and batches. Non-ideal components within the receiver, such as low-noise amplifiers, filters, and analog-to-digital converters, introduce unique hardware distortions. This "receiver fingerprint" is coupled and superimposed with the "radiation source fingerprint" in the received signal, causing a shift in the signal feature distribution of the same radiation source acquired by different receivers. This domain shift caused by receiver changes violates the requirement of deep learning models for independent and identically distributed data, resulting in a sharp decline in recognition performance when a model trained on data from a single receiver is directly deployed to a new receiver.

[0003] To mitigate the impact of receiver differences, existing research mainly focuses on two types of methods: one is based on transfer learning, which uses a small amount of labeled or unlabeled data from the target domain (new receiver) to transfer knowledge from the source domain (training receiver) to the target domain through domain adaptation or fine-tuning techniques. However, most of these methods require the acquisition of target domain data, which is limited in non-cooperative scenarios due to difficulties in data supervision or advance acquisition. The other type is based on data fusion, which trains a robust model by fusing data from multiple receivers. However, it only relies on data diversity to improve generalization ability and fails to explicitly eliminate receiver-related interference features at the feature level, thus failing to solve the distribution shift problem caused by receiver differences.

[0004] In summary, existing methods suffer from drawbacks such as strong dependence on target domain data and insufficient generalization ability across receivers, making it difficult to meet the needs of real-world scenarios where direct deployment without target domain data is not required. Therefore, developing a cross-receiver radiation source individual identification method that does not rely on target domain data, can effectively suppress the influence of receiver differences, and has strong generalization ability has become a pressing technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for identifying individual radiation sources across receivers based on multi-domain generalization. By training the identification model using only source domain samples, it achieves high-precision and robust identification of individuals from unknown radiation sources, effectively suppressing the influence of receiver differences.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0007] On one hand, the present invention provides a method for cross-receiver radiation source individual identification based on multi-domain generalization, comprising:

[0008] Acquire the I / Q signal generated by the receiver capturing and converting the wireless signal emitted by the unknown radiation source;

[0009] After preprocessing the I / Q signals, they are input into the pre-trained radiation source individual identification model. After processing by the feature extractor and classifier, the identity information of the unknown radiation source is output.

[0010] The recognition model is trained using only source domain samples.

[0011] Optionally, preprocessing the I / Q signal includes: dividing the acquired I / Q signal into I / Q signals with fixed lengths.

[0012] Optionally, the training of the recognition model includes:

[0013] The continuous I / Q signals of multiple radiation sources acquired and generated by at least two different receivers are segmented to obtain raw I / Q signal samples of fixed length. Each raw sample is labeled with the corresponding radiation source device ID to obtain a labeled source domain radiation source signal dataset.

[0014] Construct a hybrid domain adversarial alignment neural network model, which includes at least a feature extractor, a classifier, and a domain discriminator;

[0015] Perform a mixed-domain data augmentation operation on the dataset to generate mixed I / Q signal samples;

[0016] The original I / Q signal samples and the mixed samples are input into the feature extractor in the neural network model to obtain source domain features and mixed features, respectively.

[0017] Based on the source domain features and hybrid features, the joint objective function is optimized synchronously, the parameters of the neural network model are adjusted through backpropagation, and the optimal neural network model is saved as the recognition model.

[0018] Optionally, the joint objective function includes: classification loss, domain alignment loss, and adversarial loss;

[0019] Among them, domain alignment loss, adversarial loss, and classification loss are calculated for source domain features;

[0020] For mixed features, only the classification loss is calculated.

[0021] Optionally, the hybrid domain data augmentation operation is as follows: Operations, including:

[0022] Two samples were randomly selected from the source domain radiation source signal dataset. , ), ( , ),in and These represent the original I / Q signal samples respectively. and The corresponding radiation source individual label, i.e., the device ID; generate virtual samples according to the following formula ( , ):

[0023] ;

[0024] ;

[0025] in, For mixed signal samples, For the mixed soft label, Let be the interpolation coefficients, satisfying This determines which original sample the virtual sample is more biased towards; The hyperparameters used to control the interpolation intensity.

[0026] Optionally, the feature extractor is a one-dimensional convolutional neural network, including multiple sequentially connected convolutional blocks and fully connected layers connected after the convolutional blocks; each convolutional block extracts features from the original and mixed samples of the I / Q signal step by step through convolutional layers, batch normalization layers and activation function layers, and outputs a high-dimensional feature vector by the fully connected layer.

[0027] Optionally, the domain alignment loss employs multi-kernel maximum mean difference loss to align the distribution of signals from different source domains in the feature space;

[0028] The multi-kernel maximum mean difference loss is calculated by combining multiple Gaussian kernel functions with different bandwidths. For M source domains, the multi-kernel maximum mean difference loss function is defined as follows:

[0029] ;

[0030] in, Indicates feature extractor, , The first and the Data from each source domain, This represents the square of the maximum mean difference among the multi-core samples.

[0031] Optionally, the adversarial loss is generated through adversarial training between the domain discriminator and the feature extractor;

[0032] The feature extractor and the domain discriminator are connected via a gradient inversion layer. During backpropagation, the gradient flowing to the feature extractor is multiplied by a negative coefficient. ;

[0033] The optimization objective of the adversarial training is to minimize the classification loss while using a gradient reversal layer to enable the feature extractor to maximize the discrimination error of the domain discriminator, thereby allowing the feature extractor to learn domain-invariant features.

[0034] Optionally, the joint objective function is defined as: ;

[0035] in, It is the sum of the source domain feature classification loss and the mixed feature classification loss. For domain alignment loss, To combat the losses, and They are in balance and The weight.

[0036] In a second aspect, the present invention provides a computer device, comprising:

[0037] Memory, used to store computer programs;

[0038] A processor for executing the computer program to implement the steps of the cross-receiver radiation source individual identification method based on multi-domain generalization as described in any one of the first aspects.

[0039] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0040] This invention eliminates the need for target domain data in training the identification model, making it highly practical. It utilizes only readily available known receiver data for training to obtain a model with strong generalization ability to unknown receivers, thus solving the problem of identifying individuals from unknown radiation sources in data-constrained scenarios. Furthermore, this invention organically combines Mixup data augmentation, explicit domain alignment based on MK-MMD, and implicit domain-invariant feature learning based on adversarial training. Mixup smooths the data manifold and creates intermediate domains at the input level; MK-MMD brings the domains closer together at the global statistical distribution level; and adversarial training removes domain-specific features at the decision boundary level. These three elements complement each other, jointly promoting the learning of highly generalizable domain-invariant features. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the cross-receiver radiation source individual identification method based on multi-domain generalization provided in an embodiment of the present invention;

[0042] Figure 2A schematic diagram of the Hybrid Domain Adversarial Alignment (HADA) framework provided in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the feature extractor and classifier structure provided in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the domain discriminator structure provided in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram comparing the average recognition rates of various methods under different numbers of source domain receivers, as provided in the embodiments of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0047] Example 1

[0048] This embodiment introduces a method for identifying individual radiation sources across receivers based on multi-domain generalization, such as... Figure 1 As shown, it specifically includes:

[0049] It acquires the I / Q signal generated by the receiver capturing and converting the wireless signal emitted by an unknown radiation source.

[0050] The I / Q signal is segmented (using a fixed-length window for truncation) to obtain an I / Q signal with a fixed length, which is then input into a pre-trained radiation source individual identification model. After processing by a feature extractor and a classifier, the identity information of the unknown radiation source is output. The individual with the highest predicted probability is selected as the identification result of the unknown radiation source.

[0051] This embodiment focuses on the training of the identification model, which is based on an adversarial alignment framework, and its architecture is as follows: Figure 2 As shown, the specific training process is as follows:

[0052] Step S1: Training sample acquisition and preprocessing;

[0053] This embodiment uses a real-world multi-receiver radiated source signal dataset. This dataset was collected in a controlled indoor environment using three SM200B receivers and three BB60C receivers. Data from any one receiver was used for subsequent testing; training used data from all five receivers as source domain data. Signals were collected from seven wireless ad hoc networking devices (radiated sources) of the same model. The acquisition parameters were: center frequency 5.56 GHz, sampling bandwidth 27 MHz, and sampling rate 50 MSample / s. The continuous signal collected and generated by each receiver was segmented to obtain fixed-length I / Q signal samples, and each sample was labeled with its corresponding radiated source device ID, forming a labeled dataset. .

[0054] Step S2: Construct a hybrid domain adversarial alignment network model;

[0055] Feature extractor , structure as Figure 3 As shown, the input is the raw I / Q signal (two-dimensional, including real and imaginary parts). The network is a one-dimensional convolutional network containing 11 one-dimensional convolutional blocks. Each convolutional block consists of a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function in sequence. The number of channels is gradually increased to extract multi-level features. Finally, a 300-dimensional feature vector is output through a fully connected layer.

[0056] Classifier ,like Figure 3 Connected after the feature extractor, it consists of a fully connected layer and a Softmax function. The input is a 300-dimensional feature vector, and the output is the probability distribution of 7 radiation sources.

[0057] Domain discriminator , structure as Figure 4 As shown, this is a three-layer perceptron. The input is a 300-dimensional feature vector. It passes through a fully connected layer, a batch normalization layer, and a ReLU activation function. Finally, it outputs a 5-dimensional vector through the Softmax function, which corresponds to the discrimination probability of the five source domain receivers.

[0058] exist and Insert gradient inversion layers between them.

[0059] Step S3: Perform Mixup data enhancement;

[0060] During the model training phase, mixed-domain data augmentation is performed on the source domain dataset. Specifically, the Mixup technique is used to randomly linearly interpolate samples and their labels from different receiver domains to generate a series of virtual mixed-domain samples, thereby expanding the distribution range and diversity of training samples at the data level.

[0061] In each training iteration, a batch of data is randomly selected from all source domain data. Within this batch, samples are randomly paired. and and from Beta distribution Medium sampling mixing coefficient In this embodiment, the following settings are provided. =0.2. Virtual samples are generated according to the following formula ( , ):

[0062] ;

[0063] ;

[0064] Among them, and Represents the original I / Q signal sample. and This represents the individual label of the corresponding radiation source, i.e., the device ID; For mixed signal samples, For the mixed soft label, , These are the interpolation coefficients, which determine which original sample the virtual sample is more biased towards; The hyperparameters used to control the interpolation intensity.

[0065] The original batch and the mixed batch are used together for subsequent training.

[0066] Step S4: Jointly optimize and train the model;

[0067] The original and mixed samples of the I / Q signals are input into the feature extractor. The source domain features and the mixed features are obtained respectively;

[0068] During model training, the joint objective function is minimized. The joint objective function is:

[0069] ;

[0070] In this embodiment, the settings are as follows: . The classification loss (standard cross-entropy loss) for the original samples Compared to classification loss for mixed samples (Mixup cross-entropy loss) ) and. Multi-core maximum mean difference loss is employed to calculate the pairwise characteristic distribution differences between all source domain receivers. A set of bandwidth parameters is used. The Gaussian kernels are combined, and the MK-MMD² value is calculated and averaged as the alignment loss. This loss forces the mean values ​​of the feature distributions in different domains to be as close as possible in the reproducing kernel Hilbert space. To combat losses, the domain discriminator The goal is to minimize its discriminative loss to accurately distinguish which source domain a feature originates from. The gradient reversal layer, during backpropagation, will pass... gradient multiplied by This means optimization objectives and Conversely, that is, maximizing The discrimination error, thereby incentivizing Generate The characteristics of the source domain cannot be distinguished.

[0071] Among them, the standard cross-entropy loss for the original sample Measuring the probability distribution of model predictions The difference between the true label distribution y and the actual label distribution y. For a single sample, the formula is:

[0072] ;

[0073] in, It is the total number of categories; The true label of the input sample in the category The one-hot encoded value (0 or 1) on the screen; This indicates that the model predicts the category to which the input sample belongs. The probability of;

[0074] Using generated virtual samples ( , Supervise the model training process. Specifically, for each virtual sample... The model's predicted output Need to be mixed with the label Calculate the cross-entropy loss, specifically the Mixup cross-entropy loss for mixed samples. The definition is as follows:

[0075]

[0076] The multi-kernel maximum mean difference loss is used to align the distribution of signals from different source domains in the feature space. It is calculated by combining multiple Gaussian kernel functions with different bandwidths. For M source domains, the multi-kernel maximum mean difference loss function is defined as follows:

[0077] ;

[0078] in, Indicates feature extractor, , The first and the Data from each receiver domain It is the square of the maximum mean difference.

[0079] This embodiment uses a stochastic gradient descent optimizer with a learning rate of 0.01 and 100 training epochs.

[0080] Step S5: Model testing and performance evaluation;

[0081] This example adopts the "leave one domain" evaluation strategy and designs 6 different tasks (T1-T6). Each time, data from 5 receivers are selected as the source domain to train the model, and the data from the remaining receiver is used as the target domain to test the model. The specific task settings are shown in Table 1.

[0082] Table 1. Receiver settings for source and target domains in different tasks.

[0083]

[0084] Experimental Results and Analysis:

[0085] Comparison with baseline methods: The proposed method (HADA) was compared with seven baseline methods, namely Empirical Risk Minimization (ERM), Mixup, Maximum Mean Difference (MMD), Domain Adversarial Neural Network (DANN), Correlation Alignment (CORAL), Domain Invariant Feature Mining (DIFEX), and Variance Risk Extrapolation (VREx). The results are shown in Table 2. As can be seen from the table, HADA achieved an average recognition rate of 97.60% on six tasks, which is significantly higher than all other methods. The best baseline method had an average recognition rate of 95.69%, thus proving the effectiveness of the hybrid domain adversarial alignment framework.

[0086] Table 2. Performance comparison results of different domain generalization methods on cross-receiver identification tasks (%)

[0087]

[0088] Analysis of the impact of the number of source domain receivers: Figure 5 The study demonstrates the change in average recognition rate for each method as the number of source domain receivers increases from 2 to 5. The HADA method maintains its leading position in all settings, and its performance shows the largest increase with the number of source domains, indicating that this method can effectively utilize multi-source domain information to improve generalization ability.

[0089] Example 2

[0090] This embodiment provides a computer device, including:

[0091] Memory, used to store computer programs;

[0092] A processor is configured to execute the computer program to implement the steps of the cross-receiver radiation source individual identification method based on multi-domain generalization as described in any one of Embodiments 1.

[0093] In summary, this invention does not require target domain data for training the recognition model, making it highly practical. It only requires training with data from multiple readily available known receivers to obtain a model with strong generalization ability to unknown receivers, thus solving the problem of identifying individuals from unknown radiation sources in data-constrained scenarios. Furthermore, this invention organically combines Mixup data augmentation, explicit domain alignment based on MK-MMD, and implicit domain-invariant feature learning based on adversarial training. Mixup smooths the data manifold and creates intermediate domains at the input level; MK-MMD brings the domains closer together at the global statistical distribution level; and adversarial training removes domain-specific features at the decision boundary level. These three elements complement each other, jointly promoting the learning of highly generalizable domain-invariant features.

[0094] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for identifying individual radiation sources across receivers based on multi-domain generalization, characterized in that, include: Acquire the I / Q signal generated by the receiver capturing and converting the wireless signal emitted by the unknown radiation source; After preprocessing the I / Q signals, they are input into the pre-trained radiation source individual identification model. After processing by the feature extractor and classifier, the identity information of the unknown radiation source is output. The recognition model is trained using only source domain samples.

2. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 1, characterized in that, Preprocessing the I / Q signal includes dividing the acquired I / Q signal into I / Q signals with fixed lengths.

3. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 1, characterized in that, The training of the recognition model includes: The continuous I / Q signals of multiple radiation sources acquired and generated by at least two different receivers are segmented to obtain raw I / Q signal samples of fixed length. Each raw sample is labeled with the corresponding radiation source device ID to obtain a labeled source domain radiation source signal dataset. Construct a hybrid domain adversarial alignment neural network model, which includes at least a feature extractor, a classifier, and a domain discriminator; Perform a mixed-domain data augmentation operation on the dataset to generate mixed I / Q signal samples; The original I / Q signal samples and the mixed samples are input into the feature extractor in the neural network model to obtain source domain features and mixed features, respectively. Based on the source domain features and hybrid features, the joint objective function is optimized synchronously, the parameters of the neural network model are adjusted through backpropagation, and the optimal neural network model is saved as the recognition model.

4. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 3, characterized in that, The joint objective function includes: classification loss, domain alignment loss, and adversarial loss; Among them, domain alignment loss, adversarial loss, and classification loss are calculated for source domain features; For mixed features, only the classification loss is calculated.

5. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 3, characterized in that, The hybrid domain data augmentation operation is as follows: Operations, including: Two samples were randomly selected from the source domain radiation source signal dataset. , ), ( , ),in and These represent the original I / Q signal samples respectively. and The corresponding radiation source individual label, i.e., the device ID; generate virtual samples according to the following formula ( , ): ； ； in, For mixed signal samples, For the mixed soft label, Let be the interpolation coefficients, satisfying This determines which original sample the virtual sample is more biased towards; The hyperparameters used to control the interpolation intensity.

6. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 3, characterized in that, The feature extractor is a one-dimensional convolutional neural network, which includes multiple sequentially connected convolutional blocks and fully connected layers connected after the convolutional blocks. Each convolutional block extracts features from the original and mixed samples of the I / Q signal step by step through convolutional layers, batch normalization layers and activation function layers, and outputs a high-dimensional feature vector by the fully connected layer.

7. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 4, characterized in that, The domain alignment loss employs multi-kernel maximum mean difference loss to align the distribution of signals from different source domains in the feature space. The multi-kernel maximum mean difference loss is calculated by combining multiple Gaussian kernel functions with different bandwidths. For M source domains, the multi-kernel maximum mean difference loss function is defined as follows: ； in, Indicates feature extractor, , The first and the Data from each source domain, This represents the square of the maximum mean difference among the multi-core samples.

8. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 4, characterized in that, Adversarial loss is generated through adversarial training between the domain discriminator and the feature extractor; The feature extractor and the domain discriminator are connected via a gradient inversion layer. During backpropagation, the gradient flowing to the feature extractor is multiplied by a negative coefficient. ; The optimization objective of the adversarial training is to minimize the classification loss while using a gradient reversal layer to enable the feature extractor to maximize the discrimination error of the domain discriminator, thereby allowing the feature extractor to learn domain-invariant features.

9. The cross-receiver radiation source individual identification method based on multi-domain generalization according to claim 4, characterized in that, The joint objective function is defined as follows: ; in, It is the sum of the source domain feature classification loss and the mixed feature classification loss. For domain alignment loss, To combat the losses, and They are in balance and The weight.

10. A computer device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the cross-receiver radiation source individual identification method based on multi-domain generalization as described in any one of claims 1 to 9.

Images