Subdomain dynamic cross-domain radiation source identification method and device based on multi-scale features

Abstract

This invention relates to the field of wireless communication, specifically disclosing a method and apparatus for identifying cross-domain radiation sources based on multi-scale features. The method includes receiving I and Q baseband signals from a first receiver and a second receiver respectively, performing preprocessing on each signal to obtain a two-dimensional signal matrix. The two-dimensional signal matrix of the two baseband signals received by the first receiver is then labeled to obtain source domain data. The two-dimensional signal matrix of the two baseband signals received by the second receiver is set as target domain data. A cross-receiver RF fingerprinting model is constructed based on an improved ResNet18 network. The cross-receiver RF fingerprinting model is trained using the target domain data and the source domain data to obtain an optimized model. The actual target domain data is then acquired and input into the optimized model for identification, yielding the identification result. This method effectively alleviates the challenge of identifying radiation sources from cross-receivers.

Description

Technical Field

[0001] This invention relates to the field of wireless communication, and specifically to a method and apparatus for identifying subdomain dynamic cross-domain radiation sources based on multi-scale features. Background Technology

[0002] Radiation source identification utilizes non-ideal hardware features present in signals to identify different transmitters. However, when a model trained on one receiver is deployed on another, its identification performance is affected. To address the poor generalization ability of radiation source identification models across different receivers, a domain-adaptive classification method is proposed, aiming to better utilize existing data.

[0003] Existing methods for domain-adaptive radiation source identification mostly rely on discrepancy metrics or domain adversarial approaches, considering global distribution alignment but neglecting the impact of each sample on the model in different domains and the relationships between related subdomains within the same category. This leads to the loss of considerable fine-grained information. Furthermore, existing identification methods often assume that training and testing data come from the same receiver, ignoring the influence of different receiver characteristics on RF fingerprints. This is a critical issue because receivers may be replaced due to malfunction or upgrades, resulting in different receiver fingerprints that affect radiation source identification performance. Current methods for domain-adaptive radiation source identification mostly rely on discrepancy metrics or domain adversarial approaches, considering global distribution alignment but neglecting the impact of each sample on the model in different domains and the relationships between related subdomains within the same category. This leads to the loss of considerable fine-grained information. In practical applications, it is desirable for a trained model to be well-deployed on new receivers without the need for manual labeling of new data—that is, unsupervised domain adaptation. Simultaneously, it is desirable for the model to extract richer fine-grained information to improve identification performance. Summary of the Invention

[0004] To address the aforementioned problems, one objective of this invention is to provide a subdomain dynamic cross-domain radiation source identification method based on multi-scale features.

[0005] The second objective of this invention is to provide a subdomain dynamic cross-domain radiation source identification device based on multi-scale features.

[0006] The first technical solution adopted in this invention is: S1, the first receiver and the second receiver respectively receive the I and Q baseband signals, perform preprocessing on each, and obtain two-dimensional signal matrices respectively;

[0007] S2. The two-dimensional signal matrix of the two baseband signals received by the first receiver is marked with data to obtain source domain data, and the two-dimensional signal matrix of the two baseband signals received by the second receiver is set as target domain data.

[0008] S3. Based on the improved ResNet18 network, construct a cross-receiver radio frequency fingerprinting model;

[0009] S4. Using the target domain data and the source domain data, train the cross-receiver radio frequency fingerprinting model to obtain an optimized cross-receiver radio frequency fingerprinting model.

[0010] S5. Obtain the actual target domain data, input it into the optimized cross-receiver RF fingerprint recognition model for recognition, and obtain the recognition result.

[0011] Optionally, S1 includes:

[0012] S11. The first receiver and the second receiver each obtain the I and Q baseband signals through downconversion.

[0013] S12. The I and Q baseband signals are extracted using the energy detection method to obtain the steady-state signal segment;

[0014] S13. Normalize the steady-state segment of the signal to obtain a normalized signal;

[0015] S14. Based on the normalized signal, perform one-dimensional data extraction and two-dimensional data merging according to the preset number of sampling points to obtain the two-dimensional signal matrix.

[0016] Optionally, S3 includes:

[0017] S31. Use the improved ResNet18 network to extract deep features and determine the backbone network;

[0018] S32. Local feature extraction based on multi-scale convolutional neural networks;

[0019] S33. Calculate the feature distribution using the global maximum average difference and the local maximum average difference;

[0020] S34. Based on the backbone network, the local features, and the result of the feature distribution calculation, construct the cross-receiver radio frequency fingerprint recognition model.

[0021] Optionally, S4 includes:

[0022] S41. Input the target domain data into the cross-receiver radio frequency fingerprinting model, and extract deep features through the backbone network to obtain data deep features;

[0023] S42. Based on the data depth features and the source domain data, the global maximum average difference is used to calculate and determine the global distribution distance between the source domain and the target domain;

[0024] S43. Based on the data depth features, local feature extraction is performed using the multi-scale feature map region module to obtain local feature information;

[0025] S44. Using the local feature information, capture fine-grained features, and then predict the category of the target domain data to obtain the predicted label;

[0026] S45. Based on the predicted label and the source domain data, the local distribution distance between the source domain and the target domain is calculated using the local maximum average difference.

[0027] S46. Based on the local distribution distance, perform a minimization calculation to determine the overall objective optimization function;

[0028] S47. Based on the overall objective optimization function, optimize the cross-receiver radio frequency fingerprint recognition model to obtain the optimized cross-receiver radio frequency fingerprint recognition model.

[0029] Optionally, the formula for calculating the global maximum average difference is as follows:

[0030]

[0031] In the formula, MMD is a nonparametric distribution statistic, and D s For the source domain, D t For the target domain, n s and n t These represent the number of samples in the source domain data and the target domain data, respectively. and Let X be the i-th sample in the source domain and the target domain, respectively. s and X t H represents the signal sample spaces of the source and target domains, respectively. K For having a characteristic kernel K The regenerated Hilbert space;

[0032] The formula for the local maximum average difference is:

[0033]

[0034] In the formula, K represents the number of categories. and Let be the weights of the k-th transmitter in the source and target domains, respectively. Let H be the nonlinear eigenmap of the reproducing kernel Hilbert space. Is the input y i The k-th value, z, represents the output of the last fully connected layer.

[0035] Optionally, in S46,

[0036] The overall objective function is:

[0037]

[0038] In the formula, l md For globally distributed loss, l lmd For subdomain distribution loss, l ce For classification loss, λ is a non-negative trade-off factor, and α is a constant between 0 and 1.

[0039] The second technical solution adopted in this invention is: a subdomain dynamic cross-domain radiation source identification device based on multi-scale features, comprising: a preprocessing module, used to receive I and Q baseband signals through a first receiver and a second receiver respectively, perform preprocessing on each, and obtain two-dimensional signal matrices respectively;

[0040] The data marking module is used to mark the two-dimensional signal matrix of the two baseband signals received by the first receiver to obtain source domain data, and to set the two-dimensional signal matrix of the two baseband signals received by the second receiver as target domain data.

[0041] The model building module is used to build cross-receiver RF fingerprinting models based on the improved Resnet18 network.

[0042] The model training module is used to train the cross-receiver radio frequency fingerprinting model using the target domain data and the source domain data to obtain an optimized cross-receiver radio frequency fingerprinting model.

[0043] The result output module is used to acquire actual target domain data, input it into the optimized cross-receiver RF fingerprint recognition model for recognition, and obtain the recognition result.

[0044] Optionally, the preprocessing module includes:

[0045] The signal acquisition module is used for the first receiver and the second receiver to obtain the I and Q baseband signals respectively through downconversion;

[0046] The signal extraction module is used to extract the I and Q baseband signals using the energy detection method to obtain the steady-state segment of the signal;

[0047] The normalization processing module is used to normalize the steady-state segment of the signal to obtain a normalized signal;

[0048] The signal merging module is used to perform one-dimensional data extraction and two-dimensional data merging based on the normalized signal and according to a preset number of sampling points to obtain the two-dimensional signal matrix.

[0049] Optionally, the model building module includes:

[0050] A deep feature extraction module is used to extract deep features using the improved ResNet18 network and determine the backbone network.

[0051] The local feature extraction module is used to extract local features based on a multi-scale convolutional neural network.

[0052] The feature distribution calculation module is used to calculate the feature distribution using the global maximum average difference and the local maximum average difference.

[0053] The model determination module is used to construct the cross-receiver radio frequency fingerprint recognition model based on the backbone network, the local features, and the results of feature distribution calculation.

[0054] Optionally, the model training module includes:

[0055] The deep feature extraction submodule is used to input the target domain data into the cross-receiver radio frequency fingerprinting model, and perform deep feature extraction through the backbone network to obtain the data deep features.

[0056] The distribution distance calculation submodule is used to calculate and determine the global distribution distance between the source domain and the target domain based on the data depth features and the source domain data, using the global maximum average difference.

[0057] The local feature extraction submodule is used to extract local features based on the data depth features using the multi-scale feature map region module to obtain local feature information;

[0058] The prediction submodule is used to capture fine-grained features using the local feature information, and then predict the category of the target domain data to obtain the predicted label;

[0059] The local distribution distance calculation submodule is used to calculate the local distribution distance between the source domain and the target domain based on the predicted label and the source domain data, using the local maximum average difference.

[0060] The function determination submodule is used to perform minimization calculations based on the local distribution distance to determine the overall objective optimization function;

[0061] The model determination submodule is used to optimize the cross-receiver RF fingerprint recognition model based on the overall objective optimization function to obtain the optimized cross-receiver RF fingerprint recognition model.

[0062] The beneficial effects of the above technical solution are as follows:

[0063] This invention utilizes a first receiver and a second receiver to receive I and Q baseband signals respectively, perform preprocessing on each, and obtain two-dimensional signal matrices. The two-dimensional signal matrices of the two baseband signals received by the first receiver are labeled to obtain source domain data, while the two-dimensional signal matrices of the two baseband signals received by the second receiver are set as target domain data. Based on an improved ResNet18 network, a cross-receiver RF fingerprinting model is constructed. The cross-receiver RF fingerprinting model is trained using the target domain data and the source domain data to obtain an optimized cross-receiver RF fingerprinting model. Actual target domain data is then obtained and input into the optimized cross-receiver RF fingerprinting model for identification, yielding the identification result. This effectively alleviates the difficulty of identifying radiation sources from cross-receivers. Attached Figure Description

[0064] Figure 1 A flowchart of the subdomain dynamic cross-domain radiation source identification method based on multi-scale features provided by the present invention;

[0065] Figure 2 A schematic diagram of the subdomain dynamic cross-domain radiation source identification device based on multi-scale features provided by the present invention;

[0066] Figure 3 Here is a flowchart of the cross-receiver identification method according to the present invention;

[0067] Figure 4 This is a schematic diagram of the baseband signal interception method provided by the present invention;

[0068] Figure 5 This is a diagram illustrating the overall structure of the backbone network model provided by the present invention.

[0069] Figure 6 A schematic diagram of the improved ResNet18 layered structure provided for this invention;

[0070] Figure 7 A schematic diagram of the multi-scale convolutional structure in the improved ResNet18 network provided by this invention;

[0071] Figure 8 A schematic diagram of the classifier structure in the improved ResNet18 network provided by this invention. Detailed Implementation

[0072] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following detailed description of the embodiments and the accompanying drawings are used to illustrate the principles of the present invention by way of example, but should not be used to limit the scope of the present invention. That is, the present invention is not limited to the described preferred embodiments, and the scope of the present invention is defined by the claims.

[0073] In the description of this invention, it should be noted that, unless otherwise stated, "a plurality of" means two or more; the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance; those skilled in the art can understand the specific meaning of the above terms in this invention as appropriate.

[0074] Example 1

[0075] One embodiment of the present invention provides a subdomain dynamic cross-domain radiation source identification method based on multi-scale features, such as... Figure 1 As shown,

[0076] S1. The I and Q baseband signals are received by the first receiver and the second receiver respectively, and preprocessed to obtain a two-dimensional signal matrix.

[0077] Cross-receiver identification methods such as Figure 3 As shown.

[0078] S11. The first receiver and the second receiver each obtain the I and Q baseband signals through downconversion.

[0079] S12. The I and Q baseband signals are extracted using the energy detection method to obtain the steady-state signal segment;

[0080] Two receivers receive I and Q baseband signals from multiple signal transmitters and, through preprocessing, remove invalid signals and extract stable signal segments.

[0081] S13. Normalize the steady-state segment of the signal to obtain a normalized signal;

[0082] S14. Based on the normalized signal, perform one-dimensional data extraction and two-dimensional data merging according to the preset number of sampling points to obtain the two-dimensional signal matrix.

[0083] The digital receiver obtains I and Q baseband signals through down-conversion and extracts the steady-state segment of the signals using energy detection. The I and Q baseband signals are normalized, and then the data from both signals is truncated every M sampling points. Finally, the truncated one-dimensional data are merged into a two-dimensional matrix and fed into the network for recognition. In the subsequent model training phase, the I and Q two-dimensional matrices collected by the source receiver are labeled as samples, while the data in the target domain is left unlabeled. Figure 4 As shown.

[0084] S2. The two-dimensional signal matrix of the baseband signal received by the first receiver is marked with data to obtain source domain data, and the two-dimensional signal matrix of the baseband signal received by the second receiver is set as target domain data.

[0085] The data from one receiver is used as the source domain and labeled. The received signal from the other receiver is used as the target domain. Note that the target domain data does not need to be manually labeled; only the data is needed when training the model.

[0086] S3. Construct a cross-receiver radio frequency fingerprinting model using an improved ResNet18 network; specifically including the following steps S31-S34:

[0087] S31. Use the improved ResNet18 network to extract deep features and determine the backbone network;

[0088] S32. Local feature extraction based on multi-scale convolutional neural networks;

[0089] S33. Calculate the feature distribution using the global maximum average difference and the local maximum average difference;

[0090] S34. Based on the backbone network, the local features, and the result of the feature distribution calculation, construct the cross-receiver radio frequency fingerprint recognition model.

[0091] S4. Using the target domain data and the source domain data, train the cross-receiver radio frequency fingerprinting model to obtain an optimized cross-receiver radio frequency fingerprinting model.

[0092] By using known source domain data and labels, combined with target domain data, a recognition model is trained, and the recognition accuracy in the target domain is tested.

[0093] Specifically, S4 includes the following steps S41-S47:

[0094] S41. Input the target domain data into the cross-receiver radio frequency fingerprinting model, and extract deep features through the backbone network to obtain data deep features;

[0095] S42. Based on the data depth features and the source domain data, the global maximum average difference is used to calculate and determine the global distribution distance between the source domain and the target domain;

[0096] Domain-adaptive methods treat labeled training signals as the source domain and unlabeled test signals as the target domain. For the cross-receiver RF fingerprinting problem, data from the labeled receiver is used as the source domain D. s The data from another receiver that does not carry a tag is used as the target domain D. t The source dataset can be provided. Target dataset Where n s and nt X represents the number of samples in the source domain and the target domain, respectively. s and X t They represent the signal sample space, and Let Y represent the i-th sample in the source domain and the i-th sample in the target domain, respectively. S The label space is the label space for known label information in the source domain, where But Y t The label space represents the unknown label information of the target domain. In the problem studied in this paper, D s and D t Having the same label space, i.e., Y s =Y t The goal of cross-receiver RF fingerprinting is to improve recognition performance in the target domain by utilizing labeled data from the source domain and unlabeled target domain data.

[0097] The formula for calculating the global maximum average difference is as follows:

[0098]

[0099] In the formula, MMD is a nonparametric distribution statistic, and D s For the source domain, D t For the target domain, n s and n t These represent the number of samples in the source domain data and the target domain data, respectively. and Let X be the i-th sample in the source domain and the target domain, respectively. s and X t Let HK be the signal sample space of the source domain and the target domain, respectively, and let K be a region with feature kernel k. The regenerated Hilbert space;

[0100] It should be noted that MMD is a nonparametric distribution statistic used to calculate the distribution difference between the source and target domains.

[0101] S43. Based on the data depth features, local feature extraction is performed using the multi-scale feature map region module to obtain local feature information;

[0102] S44. Using the local feature information, capture fine-grained features, and then predict the category of the target domain data to obtain the predicted label;

[0103] S45. Based on the predicted label and the source domain data, the local distribution distance between the source domain and the target domain is calculated using the local maximum average difference.

[0104] To compensate for the loss of fine-grained information that global domain adaptation might miss, related subdomain adaptation can utilize local information to capture fine-grained features of each category, enabling the model to have better generalization ability and robustness during training. In practice, target domain data is often unlabeled; therefore, a multi-scale module predicts the output target domain's label, and then the extracted features and corresponding labels are input into the LMMD calculation. Furthermore, the weights represented by category information are also added to the kernel function to reduce the feature distribution of related subdomains within the same category.

[0105] The formula for the local maximum average difference is:

[0106]

[0107] In the formula, K represents the number of categories. and Let be the weights of the k-th transmitter in the source and target domains, respectively. Let H be the nonlinear eigenmap of the reproducing kernel Hilbert space. Is the input y i The k-th value, z, represents the output of the last fully connected layer.

[0108] The source domain samples are weighted by calculating the true sample labels, while the target domain is unlabeled and its weights are calculated using the predicted labels output by the network. By minimizing LLMMD, the distributions of related subdomains within the same category become more similar, thus providing fine-grained information for domain adaptation.

[0109] S46. Based on the local distribution distance, perform a minimization calculation to determine the overall objective optimization function;

[0110] like Figure 6 - Figure 8 As shown, the model's structure mainly consists of three parts: the improved ResNet18 module G... f Multi-scale module G m Label classifier G l The model's loss is composed of the globally distributed loss l. md Subdomain distribution loss l lmd And classification loss l ce Composition, calculation formula is as follows

[0111] l md =d MMD (G f (X s ),G f (X t ))

[0112] l lmd =d LMMD (G m(G f (X s )),G m (G f (X t )))

[0113]

[0114] CE(·) is the cross-entropy loss, used to train the classifier.

[0115] The overall objective function is:

[0116] m f in = m f inl ce +λ((1-α)l md +αl lmd )

[0117] In the formula, l md For globally distributed loss, l lmd For subdomain distribution loss, l ce For classification loss, λ is a non-negative trade-off factor, and α is a constant between 0 and 1.

[0118] S47. Based on the overall objective optimization function, optimize the cross-receiver radio frequency fingerprint recognition model to obtain the optimized cross-receiver radio frequency fingerprint recognition model.

[0119] Design a cross-receiver radiation source identification model, such as Figure 5 - Figure 8 As shown, firstly, an improved ResNet18 network is used as the backbone to extract deep features. Secondly, a multi-scale feature extraction module is designed, taking advantage of the wider receptive field of multi-scale networks to extract richer and finer-grained local features. Finally, maximum mean difference (MMD) and local maximum mean difference (LMMD) are used to perform global alignment and sub-domain distribution alignment of the two domains, respectively.

[0120] The model is trained and optimized by inputting known labeled data from the source domain and unknown labeled data from the target domain. During testing, the test data from the target domain is used, and the optimal network model is saved.

[0121] S5. Obtain the actual target domain data, input it into the optimized cross-receiver RF fingerprint recognition model for recognition, and obtain the recognition result.

[0122] This invention proposes a subdomain dynamic cross-domain radiation source identification method based on multi-scale features, comprising: receiving I and Q baseband signals through a first receiver and a second receiver, performing preprocessing to obtain a two-dimensional signal matrix; labeling the two-dimensional signal matrix of the baseband signal received by the first receiver to obtain source domain data; setting the two-dimensional signal matrix of the baseband signal received by the second receiver as target domain data; constructing a cross-receiver RF fingerprint identification model using an improved ResNet18 network; training the cross-receiver RF fingerprint identification model using the target domain data and the source domain data to obtain an optimized cross-receiver RF fingerprint identification model; acquiring actual target domain data; inputting it into the optimized cross-receiver RF fingerprint identification model for identification; and obtaining the identification result. Addressing the problem of cross-receiver radiation source identification, this invention proposes a subdomain dynamic adaptive cross-receiver radiation source identification model based on multi-scale features. It fully utilizes known source domain labeled data and unknown target domain data to train the designed model, enabling the model to maintain a high recognition rate in the source domain while improving the recognition accuracy in the target domain, effectively alleviating the difficulty of cross-receiver radiation source identification.

[0123] Example 2

[0124] One embodiment of the present invention provides a subdomain dynamic cross-domain radiation source identification device 200 based on multi-scale features, such as... Figure 2 As shown, it includes: a preprocessing module 201, used to receive I and Q baseband signals through the first receiver and the second receiver respectively, perform preprocessing on each, and obtain two-dimensional signal matrices respectively;

[0125] The data marking module 202 is used to mark the two-dimensional signal matrix of the two baseband signals received by the first receiver to obtain source domain data, and to set the two-dimensional signal matrix of the two baseband signals received by the second receiver as target domain data.

[0126] Model building module 203 is used to build a cross-receiver radio frequency fingerprinting model based on an improved Resnet18 network;

[0127] The model training module 204 is used to train the cross-receiver radio frequency fingerprinting model using the target domain data and the source domain data to obtain an optimized cross-receiver radio frequency fingerprinting model.

[0128] The result output module 205 is used to acquire actual target domain data, input it into the optimized cross-receiver radio frequency fingerprint recognition model for recognition, and obtain the recognition result.

[0129] Furthermore, the preprocessing module includes:

[0130] The signal acquisition module is used for the first receiver and the second receiver to obtain the I and Q baseband signals respectively through down-conversion.

[0131] The signal extraction module is used to extract the I and Q baseband signals using the energy detection method to obtain the steady-state segment of the signal;

[0132] The normalization processing module is used to normalize the steady-state segment of the signal to obtain a normalized signal;

[0133] The signal merging module is used to perform one-dimensional data extraction and two-dimensional data merging based on the normalized signal and according to a preset number of sampling points to obtain the two-dimensional signal matrix.

[0134] Furthermore, the model building module includes:

[0135] A deep feature extraction module is used to extract deep features using the improved ResNet18 network and determine the backbone network.

[0136] The local feature extraction module is used to extract local features based on a multi-scale convolutional neural network.

[0137] The feature distribution calculation module is used to calculate the feature distribution using the global maximum average difference and the local maximum average difference.

[0138] The model determination module is used to construct the cross-receiver radio frequency fingerprint recognition model based on the backbone network, the local features, and the results of feature distribution calculation.

[0139] Furthermore, the model training module includes:

[0140] The deep feature extraction submodule is used to input the target domain data into the cross-receiver radio frequency fingerprinting model, and perform deep feature extraction through the backbone network to obtain the data deep features.

[0141] The distribution distance calculation submodule is used to calculate and determine the global distribution distance between the source domain and the target domain based on the data depth features and the source domain data, using the global maximum average difference.

[0142] The local feature extraction submodule is used to extract local features based on the data depth features using the multi-scale feature map region module to obtain local feature information;

[0143] The prediction submodule is used to capture fine-grained features using the local feature information, and then predict the category of the target domain data to obtain the predicted label;

[0144] The local distribution distance calculation submodule is used to calculate the local distribution distance between the source domain and the target domain based on the predicted label and the source domain data, using the local maximum average difference.

[0145] The function determination submodule is used to perform minimization calculations based on the local distribution distance to determine the overall objective optimization function;

[0146] The model determination submodule is used to optimize the cross-receiver RF fingerprint recognition model based on the overall objective optimization function to obtain the optimized cross-receiver RF fingerprint recognition model.

[0147] This invention provides a subdomain dynamic cross-domain radiation source identification device based on multi-scale features, comprising: a preprocessing module for receiving I and Q baseband signals through a first receiver and a second receiver, performing preprocessing to obtain a two-dimensional signal matrix; a data labeling module for labeling the two-dimensional signal matrix of the baseband signal received by the first receiver to obtain source domain data, and setting the two-dimensional signal matrix of the baseband signal received by the second receiver as target domain data; a model building module for constructing a cross-receiver radio frequency fingerprint identification model using an improved ResNet18 network; a model training module for training the cross-receiver radio frequency fingerprint identification model using the target domain data and the source domain data to obtain an optimized cross-receiver radio frequency fingerprint identification model; and a result output module for acquiring actual target domain data, inputting it into the optimized cross-receiver radio frequency fingerprint identification model for identification, and obtaining the identification result. To address the problem of cross-receiver radiation source identification, this invention proposes a subdomain dynamic adaptive cross-receiver radiation source identification model based on multi-scale features. The model is trained by making full use of known source domain labeled data and unknown target domain data, which enables the model to maintain a high recognition rate in the source domain while improving the recognition accuracy in the target domain, effectively alleviating the difficulty of cross-receiver radiation source identification.

[0148] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying subdomain dynamic cross-domain radiation sources based on multi-scale features, characterized in that, include: S1. The first receiver and the second receiver respectively receive the I and Q baseband signals, perform preprocessing on each, and obtain two-dimensional signal matrices respectively. S2. The two-dimensional signal matrix of the two baseband signals received by the first receiver is marked with data to obtain source domain data, and the two-dimensional signal matrix of the two baseband signals received by the second receiver is set as target domain data. S3. Based on the improved ResNet18 network, construct a cross-receiver radio frequency fingerprinting model; including: S31. Use the improved ResNet18 network to extract deep features and determine the backbone network; S32. Local feature extraction based on multi-scale convolutional neural networks; S33. Calculate the feature distribution using the global maximum average difference and the local maximum average difference; S34. Based on the backbone network, the local features, and the result of the feature distribution calculation, construct the cross-receiver radio frequency fingerprint recognition model; S4. Using the target domain data and the source domain data, train the cross-receiver RF fingerprinting model to obtain an optimized cross-receiver RF fingerprinting model; including: S41. Input the target domain data into the cross-receiver radio frequency fingerprinting model, and extract deep features through the backbone network to obtain data deep features; S42. Based on the data depth features and the source domain data, the global maximum average difference is used to calculate and determine the global distribution distance between the source domain and the target domain; S43. Based on the data depth features, local feature extraction is performed using the multi-scale feature map region module to obtain local feature information; S44. Using the local feature information, capture fine-grained features, and then predict the category of the target domain data to obtain the predicted label; S45. Based on the predicted label and the source domain data, the local distribution distance between the source domain and the target domain is calculated using the local maximum average difference. S46. Based on the local distribution distance, perform a minimization calculation to determine the overall objective optimization function; S47. Based on the overall objective optimization function, optimize the cross-receiver radio frequency fingerprint recognition model to obtain the optimized cross-receiver radio frequency fingerprint recognition model; The formula for calculating the global maximum average difference is as follows: In the formula, MMD is a nonparametric distribution statistic, and D s For the source domain, D t For the target domain, n s and n t These represent the number of samples in the source domain data and the target domain data, respectively. and Let X be the i-th sample in the source domain and the target domain, respectively. s and X t H represents the signal sample spaces of the source and target domains, respectively. K For having a characteristic kernel K The regenerated Hilbert space; The formula for calculating the local maximum average difference is as follows: In the formula, K represents the number of categories. and Let be the weights of the k-th transmitter in the source and target domains, respectively. Let H be the nonlinear eigenmap of the reproducing kernel Hilbert space. It is input The k-th value, z, represents the output of the last fully connected layer; S5. Obtain the actual target domain data, input it into the optimized cross-receiver RF fingerprint recognition model for recognition, and obtain the recognition result.

2. The subdomain dynamic cross-domain radiation source identification method based on multi-scale features according to claim 1, characterized in that, S1 includes: S11. The first receiver and the second receiver each obtain the I and Q baseband signals through downconversion. S12. The I and Q baseband signals are extracted using the energy detection method to obtain the steady-state signal segment; S13. Normalize the steady-state segment of the signal to obtain a normalized signal; S14. Based on the normalized signal, perform one-dimensional data extraction and two-dimensional data merging according to the preset number of sampling points to obtain the two-dimensional signal matrix.

3. The subdomain dynamic cross-domain radiation source identification method based on multi-scale features according to claim 1, characterized in that, In S46, The overall objective function is: m f in=m f inl ce +λ((1-α)l md +αl lmd ) In the formula, l md For globally distributed loss, l lmd For subdomain distribution loss, l ce For classification loss, λ is a non-negative trade-off factor, and α is a constant between 0 and 1.

4. A subdomain dynamic cross-domain radiation source identification device based on multi-scale features, characterized in that, include: The preprocessing module is used to receive the I and Q baseband signals from the first receiver and the second receiver respectively, perform preprocessing on each signal, and obtain two-dimensional signal matrices respectively. The data marking module is used to mark the two-dimensional signal matrix of the two baseband signals received by the first receiver to obtain source domain data, and to set the two-dimensional signal matrix of the two baseband signals received by the second receiver as target domain data. The model building module is used to build cross-receiver RF fingerprinting models based on the improved Resnet18 network. The model building module includes: A deep feature extraction module is used to extract deep features using the improved ResNet18 network and determine the backbone network. The local feature extraction module is used to extract local features based on a multi-scale convolutional neural network. The feature distribution calculation module is used to calculate the feature distribution using the global maximum average difference and the local maximum average difference. The model determination module is used to construct the cross-receiver radio frequency fingerprinting model based on the backbone network, the local features, and the results calculated from the feature distribution. The model training module is used to train the cross-receiver radio frequency fingerprinting model using the target domain data and the source domain data to obtain an optimized cross-receiver radio frequency fingerprinting model. The model training module includes: The deep feature extraction submodule is used to input the target domain data into the cross-receiver radio frequency fingerprinting model, and perform deep feature extraction through the backbone network to obtain the data deep features. The distribution distance calculation submodule is used to calculate and determine the global distribution distance between the source domain and the target domain based on the data depth features and the source domain data, using the global maximum average difference. The local feature extraction submodule is used to extract local features based on the data depth features using the multi-scale feature map region module to obtain local feature information; The prediction submodule is used to capture fine-grained features using the local feature information, and then predict the category of the target domain data to obtain the predicted label; The local distribution distance calculation submodule is used to calculate the local distribution distance between the source domain and the target domain based on the predicted label and the source domain data, using the local maximum average difference. The function determination submodule is used to perform minimization calculations based on the local distribution distance to determine the overall objective optimization function; The model determination submodule is used to optimize the cross-receiver RF fingerprint recognition model based on the overall objective optimization function to obtain the optimized cross-receiver RF fingerprint recognition model. The formula for calculating the global maximum average difference is as follows: In the formula, MMD is a nonparametric distribution statistic, and D s For the source domain, D t For the target domain, n s and n t These represent the number of samples in the source domain data and the target domain data, respectively. and Let X be the i-th sample in the source domain and the target domain, respectively. s and X t H represents the signal sample spaces of the source and target domains, respectively. K For having a characteristic kernel K The regenerated Hilbert space; The formula for calculating the local maximum average difference is as follows: In the formula, K represents the number of categories. and Let be the weights of the k-th transmitter in the source and target domains, respectively. Let H be the nonlinear eigenmap of the reproducing kernel Hilbert space. Is the input y i The k-th value, z, represents the output of the last fully connected layer; The result output module is used to acquire actual target domain data, input it into the optimized cross-receiver RF fingerprint recognition model for recognition, and obtain the recognition result.

5. The subdomain dynamic cross-domain radiation source identification device based on multi-scale features according to claim 4, characterized in that, The preprocessing module includes: The signal acquisition module is used for the first receiver and the second receiver to obtain the I and Q baseband signals respectively through downconversion; The signal extraction module is used to extract the I and Q baseband signals using the energy detection method to obtain the steady-state segment of the signal; The normalization processing module is used to normalize the steady-state segment of the signal to obtain a normalized signal; The signal merging module is used to perform one-dimensional data extraction and two-dimensional data merging based on the normalized signal and according to a preset number of sampling points to obtain the two-dimensional signal matrix.

Images