Radiation source individual identification method based on deep sub-domain adaptation

Abstract

The present invention discloses a method for identifying individual radiation sources based on deep sub-domain adaptation, which belongs to the field of individual radiation source identification. The present invention includes obtaining source domain signals and target domain signals of the radiation source ADS-B signal, which are of one-dimensional point data; converting the one-dimensional point data into an IQ waveform diagram to form an I / Q splicing diagram; extracting common features of the source domain and target domain data; embedding a local maximum mean difference adaptive layer to align the sub-domain distribution; constructing a sub-domain adaptive model for individual radiation source identification through the local maximum mean difference and a pre-trained Resnet-50 model; training the sub-domain adaptive model and using the model probability prediction result as a pseudo label of the target domain; and using the constructed sub-domain adaptive network to perform migration identification on radiation source signals with different signal-to-noise distributions. The present invention can accurately identify individual ADS-B radiation sources under different signal-to-noise distributions while saving training time.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source individual identification, and more specifically, to a radiation source individual identification method based on deep sub-domain adaptation. Background Art

[0002] With the continuous development of wireless communications and the Internet of Things (IoT), the number of wireless devices in the spectrum is rapidly increasing. The subtle characteristics generated by wireless devices due to device tolerances have physical characteristics that are difficult to clone. These subtle characteristics that vary from device to device are called RF fingerprints. Using RF fingerprints to distinguish legitimate devices from illicit ones is a new physical layer method for protecting communication system security.

[0003] Specific Emitter Identification (SEI) is the process of extracting radio frequency fingerprints from signals from similar emitters to identify individual emitters. SEI technology has broad application prospects in both military and civilian fields. In civilian applications, SEI technology can be applied to spectrum resource management, wireless network security, and cognitive radio. In military communications, SEI technology can identify the signals of specific emitters in complex battlefield environments, which is crucial for identifying friend or foe and understanding battlefield situational awareness.

[0004] In recent years, deep learning technology has demonstrated its tremendous potential in SEI applications. When the transmission environment and receiver of the signal from the radiator to be identified are identical to those of the labeled signal used for training, neural networks, leveraging their powerful nonlinear mapping capabilities, can extract and abstract RF fingerprint characteristics, achieving excellent classification performance. However, in real-world communication scenarios, due to variations in the transmission environment and internal receiver noise, the signal-to-noise ratio (SNR) of the received signal from the same radiator may vary. When the SNR of the labeled training signal differs from that of the unlabeled test signal, the recognition performance of supervised deep learning models is significantly compromised. Although the SNRs of the training and test signals may differ, the fingerprint information they carry is consistent. Therefore, learning this "invariant" fingerprint feature information is crucial to addressing the mismatch between the training and test data distributions. Summary of the Invention

[0005] 1. Technical problem to be solved by the invention

[0006] The present invention provides a method for individual radiation source identification based on deep sub-domain adaptation, aiming to solve the problem of performance degradation of the recognition model caused by the inconsistency between the radiation source signal to be identified and the channel environment noise of the training data set.

[0007] 2. Technical solution

[0008] In order to achieve the above object, the technical solution provided by the present invention is:

[0009] The present invention provides a method for identifying individual radiation sources based on deep sub-domain adaptation, comprising:

[0010] S100, obtaining one-dimensional point data of an ADS-B signal of a radiation source, wherein the data is divided into source domain data and target domain data;

[0011] S200, converting the one-dimensional point data into an IQ waveform diagram, and splicing the I-channel image and the corresponding Q-channel image together in a splicing manner to form an I / Q splicing diagram;

[0012] S300, uses the improved image set pre-trained Resnet-50 network structure to extract common features of source domain and target domain data;

[0013] S400, embedding the local maximum mean difference adaptation layer to align the sub-domain distribution, reducing the distribution difference between the source domain and the target domain while capturing the fine information of close sub-categories;

[0014] S500: Build a subdomain adaptive model for individual radiation source identification by using the local maximum mean difference and a pre-trained ResNet-50 model. Train the subdomain adaptive model and use the model's probability prediction results as pseudo labels for the target domain. Learn the classification features shared by the source and target domains by minimizing the loss objective function.

[0015] S600: Use the constructed sub-domain adaptive network to perform migration recognition on radiation source signals with different signal-to-noise distributions.

[0016] Furthermore, step S200 normalizes the data before converting the one-dimensional point data into an IQ waveform. The normalization method includes normalizing the I-channel data and the Q-channel data of the radiation source respectively, and changing the maximum value to 1. The normalization formula is:

[0017]

[0018] Where Y is the normalized result and X is the data to be normalized.

[0019] Furthermore, the conversion of the one-dimensional point data into an IQ waveform diagram includes: taking m data points as a data sample, changing the dimension of the one-dimensional IQ data according to the horizontal axis being time and the vertical axis being the size of the sample value, and obtaining a two-dimensional waveform diagram of the IQ two-channel data.

[0020] Furthermore, the I-channel image and the corresponding Q-channel image are stitched together to form an I / Q stitching image, including:

[0021] The I-channel and Q-channel two-dimensional waveforms are spliced ​​together, with the I-channel image on the left and the Q-channel image on the right.

[0022] Furthermore, considering the additive white Gaussian noise channel scenario for the IQ spliced ​​image, forming an IQ spliced ​​image with different noise distributions includes:

[0023] The signal-to-noise ratio range is 0 to kdB, and there are k / 2 IQ splicing datasets with different noise environments at 2dB intervals. The dataset with a signal-to-noise ratio of n dB is taken as the source domain data, and the remaining k / 2-1 noise datasets are the target domain datasets.

[0024] Furthermore, the pre-trained Resnet-50 model described in step S300 is a 50-layer residual neural network model trained on millions of images from the ImageNet dataset. It includes 49 convolutional layers and one fully connected layer with 1,000 nodes. All convolutional layers are used as feature extractors, the last fully connected layer is removed, and a two-layer fully connected layer is added. The first fully connected layer of the network is set as the adaptive layer, and sub-domain adaptation is performed using the local maximum mean difference criterion.

[0025] Furthermore, the local maximum mean difference adaptation layer is improved based on the maximum mean difference that measures the overall difference in the distribution between the source domain and the target domain data. By introducing the local maximum mean difference to perform conditional distribution difference calculation, it is minimized during the iterative process of the deep network, thereby narrowing the distribution differences of related subdomains within the same category.

[0026] Furthermore, the local maximum mean difference expression is:

[0027]

[0028] in, and They are and The weight belonging to category c is calculated as:

[0029]

[0030] Furthermore, for samples in the source domain, the true labels are used Calculate the weight of each sample as one-hot For unlabeled target domains, deep neural networks The output is a probability distribution, using the network to predict the label As the pseudo label of the target domain to calculate the weight of the target domain

[0031] To achieve self-adaptation of high-level network layers, we need to know the activation value z l ; Given n s A labeled source domain D s , n t There is an unlabeled target domain D t , respectively obey the probability p and q, the deep neural network will produce in the l layer and The activation value of ; the sub-domain adaptive function is defined as:

[0032] Where Z l is the feature extracted from the lth layer, l∈L={1,2,…,|L|}.

[0033] During the model training process, LMMD in the above formula is used for domain adaptation loss, so that the objective function of network optimization is:

[0034]

[0035] Where J(·,·) is the cross entropy loss function; is the sub-domain adaptive function; Le is the total number of adaptive layers.

[0036] Furthermore, the sub-domain adaptive model is trained using the I / Q splicing sample data of the source domain radiation source signal with a signal-to-noise ratio of n. The I / Q splicing graphs of the radiation sources with k / 2-1 signal-to-noise ratios to be identified are input into the trained sub-domain adaptive model to perform transfer recognition on the target domain radiation source individuals.

[0037] 3. Beneficial effects

[0038] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0039] The present invention's method for identifying individual radiation sources based on deep sub-domain adaptation uses the pre-trained deep model Resnet-50 on an image set and performs fine-tuning. Compared with the traditional fully trained neural network model, it can better extract the radiation source features under different noise conditions; embeds a local maximum mean difference adaptive layer to align the sub-domain distribution, while reducing the distribution difference between the source domain and the target domain, and improves the model recognition performance by capturing the fine information of close-range categories. In response to the situation where the target domain has no label, the trained network model probability prediction result is used to obtain the pseudo label of the target domain for solution, which can more accurately identify the type of individual radiation sources under different noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the method flow of the present invention;

[0041] Figure 2 Schematic diagram of I-path and Q-path of two types of radiation sources in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of image stitching in an embodiment of the present invention. Figure 3 (a) is the schematic diagram of I / Q splicing in sample 1. Figure 3 (b) is a schematic diagram of I / Q splicing in sample 2;

[0043] Figure 4 Schematic diagram of IQ splicing under different signal-to-noise ratios in an embodiment of the present invention. Figure 4 (a) is the I / Q splicing diagram of sample 1 at a signal-to-noise ratio of 0dB. Figure 4 (b) is the I / Q splicing diagram of sample 1 at a signal-to-noise ratio of 10 dB;

[0044] Figure 5 Schematic diagram showing the difference between global domain adaptation and sub-domain adaptation in an embodiment of the present invention;

[0045] Figure 6 Schematic diagram of the sub-domain adaptive layer feature alignment structure DSAN in an embodiment of the present invention;

[0046] Figure 7 : is a schematic diagram of the confusion matrix of the model before and after optimization of the transfer learning method in an embodiment of the present invention, Figure 7 (a) is a schematic diagram of the Resnet50 confusion matrix. Figure 7 (b) is a schematic diagram of the DSAN confusion matrix;

[0047] Figure 8 Schematic diagram of training loss of different models in embodiments of the present invention;

[0048] Figure 9 It is a schematic diagram of feature visualization of different models in an embodiment of the present invention when migrating from a source domain signal-to-noise ratio of 10 dB to a target domain signal-to-noise ratio of 14 dB. DETAILED DESCRIPTION

[0049] In order to further understand the content of the present invention, the present invention is described in detail with reference to the accompanying drawings and embodiments.

[0050] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for understanding and reading by those familiar with this technology. They are not used to limit the conditions for implementation of the present invention and therefore have no substantial technical significance. Any modification of the structure, change in the proportional relationship, or adjustment of the size should still fall within the scope of the technical content disclosed by the present invention without affecting the efficacy and purpose of the present invention. At the same time, terms such as "upper", "lower", "left", "right", and "middle" quoted in this specification are only for the convenience of description and are not used to limit the scope of implementation. Changes or adjustments in their relative relationships should also be considered as the scope of implementation of the present invention without substantially changing the technical content.

[0051] The present invention provides a method for identifying individual radiation sources based on deep sub-domain adaptation, comprising the following steps:

[0052] S100, obtaining one-dimensional point data of an ADS-B signal of a radiation source, wherein the data is divided into source domain data and target domain data;

[0053] S200, converting the one-dimensional point data into an IQ waveform diagram, and splicing the I-channel image and the corresponding Q-channel image together in a splicing manner to form an I / Q splicing diagram;

[0054] S300, uses the improved image set pre-trained Resnet-50 network structure to extract common features of source domain and target domain data;

[0055] S400, embedding the local maximum mean difference adaptation layer to align the sub-domain distribution, reducing the distribution difference between the source domain and the target domain while capturing the fine information of close sub-categories;

[0056] S500: Build a subdomain adaptive model for individual radiation source identification by using the local maximum mean difference and a pre-trained ResNet-50 model. Train the subdomain adaptive model and use the model's probability prediction results as pseudo labels for the target domain. Learn the classification features shared by the source and target domains by minimizing the loss objective function.

[0057] S600: Use the constructed sub-domain adaptive network to perform migration recognition on radiation source signals with different signal-to-noise distributions.

[0058] Specifically, in step S100 , one-dimensional point data of the ADS-B signal of the radiation source is obtained.

[0059] Furthermore, before converting the one-dimensional point data into an IQ waveform, the data needs to be normalized. The normalization method includes normalizing the I-channel data and the Q-channel data of the radiation source respectively, and changing the maximum value to 1. The normalization formula is:

[0060]

[0061] Where Y is the normalized result and X is the data to be normalized.

[0062] In step S200, the one-dimensional point data is converted into an IQ waveform graph, and the I-channel image and the corresponding Q-channel image are spliced ​​together using a splicing method to form an I / Q spliced ​​graph. The conversion of the one-dimensional point data into the IQ waveform graph includes taking m data points as a data sample, transforming the one-dimensional IQ data with time as the horizontal axis and sample value as the vertical axis, to obtain a two-dimensional waveform graph of the IQ data. The I-channel image and the corresponding Q-channel image are then spliced ​​together using a splicing method to form an I / Q spliced ​​graph, including splicing the two-dimensional I and Q waveform graphs, with the I-channel image on the left and the Q-channel image on the right.

[0063] In step S200, an additive white Gaussian noise (AWGN) channel scenario is considered for the IQ stitched image. This process can form an IQ stitched image with different noise distributions, including:

[0064] The signal-to-noise ratio range is 0 to kdB, and there are k / 2 IQ splicing datasets with different noise environments at 2dB intervals. The dataset with a signal-to-noise ratio of n dB is taken as the source domain data, and the remaining k / 2-1 noise datasets are the target domain datasets.

[0065] In step S500, a subdomain adaptive model (DSAN) for individual radiator identification is constructed using a pretrained ResNet-50 model and a local maximum mean difference (LMMD) adaptive layer. The DSAN model includes a ResNet-50 model, a general feature extractor pretrained on the image set, a LMMD adaptive layer, and a fully connected layer.

[0066] The pre-trained Resnet-50 model is a 50-layer residual neural network model trained on millions of images from the ImageNet dataset. It consists of 49 convolutional layers and one fully connected layer with 1,000 nodes. This paper uses all convolutional layers as feature extractors, removes the last fully connected layer, and adds a two-layer fully connected layer. The first fully connected layer is set as the adaptive layer, and sub-domain adaptation is performed using the local maximum mean difference criterion.

[0067] The local maximum mean discrepancy criterion (LMMD) is expressed as:

[0068]

[0069] in, and They are and The weight belonging to category c is calculated as:

[0070]

[0071] Where: y ic is the input vector Y i Category c label.

[0072] Input the source domain data of the radiation source signal to train the DSAN model, and use the output of the network model as the pseudo label of the target domain data According to the source domain label Y s and pseudo labels Find the sample category weight Calculate the LMMD value of the source domain and target domain training network of the feature adaptation layer

[0073] For samples in the source domain, use the true label Calculate the weight of each sample as one-hot For unlabeled target domains, deep neural networks The output of is a probability distribution that well characterizes the i The probability of being identified as class C is determined by the network prediction label. As the pseudo label of the target domain to calculate the weight of the target domain

[0074] To achieve self-adaptation of high-level network layers, we need to know the activation value z l Given a s A labeled source domain D s , n t There is an unlabeled target domain D t , respectively obey the probability p and q, the deep neural network will produce in the l layer and Therefore, the subdomain adaptation function is:

[0075] Where Z l is the feature extracted from the lth layer (l∈L={1,2,…,|L|}).

[0076] During the model training process, LMMD is used for domain adaptive loss to obtain the objective function of network optimization; the loss objective function L is minimized until the model converges or the number of cycles reaches the maximum number of training rounds E. The loss objective function expression is:

[0077]

[0078] Where J(·,·) is the cross entropy loss function; is the sub-domain adaptive function; Le is the total number of adaptive layers.

[0079] Finally, the sub-domain adaptive network DSAN trained with the radiation source signal data with a signal-to-noise ratio of n is used to perform migration recognition on the unlabeled radiation source signals with k / 2-1 signal-to-noise ratios to obtain the signal classification and recognition results.

[0080] For further analysis, we sampled 4800 points of the normalized IQ data, for a total of 500 samples per category. Finally, we converted the one-dimensional data points into a two-dimensional waveform graph, with the horizontal axis representing the time series and the vertical axis representing the corresponding values ​​of the one-dimensional sequence. Figure 2 The following diagram shows some IQ data waveforms of two types of radiation source signals.

[0081] Through the observation of IQ data, it is found that the I-channel data and Q-channel data will change accordingly. There is repetitiveness in the data with the same label and difference in the data with different labels. In order to better take into account and integrate the complete characteristics of the I-channel and Q-channel signals, the I and Q-channel signal values-sampling points are spliced ​​into an image to form an I / Q splicing graph. The splicing graph of the first data in the first category is as follows: Figure 3 As shown in (a), the splicing diagram of the first data in the second category is as follows Figure 3 (b) shown.

[0082] Consider the Additive White Gaussian Noise (AWGN) channel scenario. Figure 4 The IQ mosaic images of the first type of radiation source signal with a signal-to-noise ratio of 0dB and 10dB are given. It can be seen that due to the difference in signal-to-noise ratio, the two mosaic images present a great visual difference. It is difficult to obtain unified features using a model trained with samples under a single noise condition.

[0083] Domain adaptation mainly eliminates domain differences by mapping the source domain and the target domain into a common feature space and re-forms a feature set with the same distribution. It can be divided into two parts: global domain adaptation and sub-domain adaptation, such as Figure 5 As shown in Figure 2. Global domain adaptation mainly learns global domain movement through the network model, that is, aligning the global source domain and target domain distribution without considering the relationship between the same sub-radiation source categories in the two global domains, resulting in a small distribution distance between sub-domains and causing misidentification, such as Figure 5 Shown in the left half. Figure 5The right half represents subdomain adaptation (SDA). When the subdomain distributions of the source and target domains are aligned, the global distribution is also roughly the same, which can significantly improve classification accuracy. Therefore, using subdomain adaptation in environments with different signal-to-noise ratios can better match the distribution differences of individual radiant categories in the source and target domains.

[0084] like Figure 6 The figure shows the network structure of DSAN. During network training, the convolutional layer parameters are shared, but in the fully connected layer, the alignment of the fully connected layer feature parameters is achieved by embedding the domain adaptation unit LMMD.

[0085] In order to compare the effects of the present invention, corresponding comparison models are constructed. Model 1 Resnet50 is a TCNN network structure, which uses the image set pre-trained ReNet-50 fine-tuning transfer learning method and adds three fully connected layers after the global average pooling layer; Model 2 DeepCoral, based on Model 1, applies CORAL loss to the last fully connected layer; Model 3 DAN, based on Model 1, embeds multi-core MMD measurement in the second last fully connected layer; Model 4 MRAN, replaces the measurement criterion with CMMD; Model 5 DANN, draws on the idea of ​​generating adversarial networks to add a domain discriminator after the feature extractor of Model 1, and connects them through a gradient reversal layer (GRL) in the middle; Model 6 is the method of this article, the general feature extraction layer is the image set pre-trained ResNet-50, and LMMD is used for the loss of the domain adaptation layer. The parameters of the adaptive layer are the same as the second last fully connected layer in Model 1, and the fully connected layer is the same as the last fully connected layer. The recognition accuracy of different transfer learning models is shown in Table 1. The transfer task A→B indicates that the data in case A is the source domain, and the data in case B is the target domain. This paper uses labeled radiation source signals with a signal-to-noise ratio of 10dB as the source domain, and unlabeled signals at other signal-to-noise ratios as the target domain. The ratio of source domain data to target domain data is set to 5:1, that is, the source domain data is 5000 samples with a signal-to-noise ratio of 10dB, and the target domain data is 1000 samples with signal-to-noise ratios of n (n = 0, 2, 4, 6, 8, 10, 12, 14).

[0086] Table 1 Recognition accuracy of different transfer learning models

[0087] Migration tasks 10→0 10→2 10→4 10→6 10→8 10→10 10→12 10→14 Resnet50 46.8 54.4 62.7 72.2 76.2 88.7 76.7 79.3 DeepCoral 61.1 67.2 72.9 78.9 82.9 95.6 82.2 83.4 DAN 61.1 67.0 73.6 78.8 81.4 94.7 83.3 85.8 MRAN 64.0 70.9 79.7 83.9 85.1 96.2 87.8 87.2 DANN 62.3 70.3 76.3 83.8. 85.9 97.0 89.4. 88.1 DSAN 65.6. 73.0 79.8 84.9 86.5 98.0 90.2 90.7

[0088] Table 1 shows that the DSAN model achieved the highest accuracy in all transfer tasks under eight different channel noise environments, demonstrating the proposed method's superior generalization and robustness. A comparison of Model 2 and Model 1 reveals that the fine-tuning of the ResNet-50 pre-trained on the image set yields limited fingerprint features. Domain adaptation, which aligns the source and target domains, shortens the marginal distribution distance and exhibits better feature extraction under varying data distributions. Models 2 through 5 all achieved improved recognition accuracy compared to the fine-tuned ResNet-50 model, but their generalization was limited due to the lack of subdomain adaptation. The proposed model leverages the excellent general feature extraction capabilities of the deep neural network ResNet-50 and employs a subdomain adaptation metric to align features of different subclasses of radiators, demonstrating superior unsupervised recognition across signal-to-noise ratios. Model 6 and Model 3 demonstrate that aligning subdomains of the source and target domains yields superior results compared to global domain adaptation, further improving both generalization and robustness.

[0089] In order to compare the classification and recognition effects of various radiation source signals after using the transfer learning method, the signal-to-noise ratio in the source domain is 10dB and the signal-to-noise ratio in the target domain is 14dB. Figure 7 The confusion matrices obtained after training using only the fine-tuned ResNet50 model and optimizing it using the subdomain adaptation method are presented. Before the algorithm optimization, the model's confusion matrix exhibited significant errors, with only the second and fifth categories of radiation source signals achieving classification accuracy exceeding 90%. After the optimization, the confusion matrix significantly approached the identity matrix, with only the sixth, ninth, and tenth categories achieving accuracy below 90%. This demonstrates the significant improvement in classification results achieved using the subdomain adaptation method.

[0090] To verify the convergence of the invented model, the convergence of DSAN, DANN and DAN is proved. Taking the source domain to target domain 10→14 as an example, Figure 8 The figure shows the training loss of different models after 100 training rounds. At the same number of iterations, DSAN achieved lower training loss and faster convergence than both the domain adaptation model DAN (using the multi-core MMD metric) and the domain adaptation model DANN (using adversarial training). However, DANN exhibited significant fluctuations, and its training loss did not converge to its minimum value.

[0091] To further understand the gain of the algorithm of the present invention, the output of the first fully connected layer of the model before and after optimization is reduced to two dimensions using the t-SNE algorithm and visualized. The input is the IQ splicing graph after signal preprocessing, the source domain signal-to-noise ratio is 10dB, and the target domain signal-to-noise ratio is 14dB. The visualization results of the domain adaptation model DAN using the multi-core MMD metric and the sub-domain adaptation model DSAN used in this paper are shown in the figure below. Figure 9 As shown in the figure. After the optimization, the extracted target domain features are more separable, as evidenced by a tighter distribution of features for signals from the same source and clearer boundaries for features from different sources. This is because the proposed method increases the distinguishability of feature samples by reducing the distance between the subdomains of the source and target domains. By combining the advantages of ResNet-50 and subdomain adaptation, "invariant" fingerprint features from signals with different signal-to-noise ratios in the source and target domains are more tightly clustered.

[0092] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A method for identifying individual radiation sources based on deep sub-domain adaptation, characterized in that: include: S100, obtaining one-dimensional point data of an ADS-B signal of a radiation source, wherein the data is divided into source domain data and target domain data; S200, converting the one-dimensional point data into an IQ waveform diagram, and splicing the I-channel image and the corresponding Q-channel image together in a splicing manner to form an I / Q splicing diagram; S300, using an improved image set pre-trained Resnet-50 network structure to extract common features of source domain and target domain data; the pre-trained Resnet-50 model is a 50-layer residual neural network model trained on millions of images in the ImageNet dataset, which includes 49 convolutional layers and a 1000-node fully connected layer; all convolutional layers are used as feature extractors, the last fully connected layer is removed and a two-layer fully connected layer is added, and the first fully connected layer network is set as an adaptive layer, and sub-domain adaptation is performed using the local maximum mean difference criterion; S400, embedding the local maximum mean difference adaptation layer to align the sub-domain distribution, reducing the distribution difference between the source domain and the target domain while capturing the fine information of close sub-categories; S500: Build a subdomain adaptive model for individual radiation source identification by using the local maximum mean difference and a pre-trained ResNet-50 model. Train the subdomain adaptive model and use the model's probability prediction results as pseudo labels for the target domain. Learn the classification features shared by the source and target domains by minimizing the loss objective function. S600: Use the constructed sub-domain adaptive network to perform migration recognition on radiation source signals with different signal-to-noise distributions.

2. The radiation source individual identification method based on deep sub-domain adaptation according to claim 1 is characterized in that: Step S200 normalizes the data before converting the one-dimensional point data into an IQ waveform. The normalization method includes normalizing the I-channel data and the Q-channel data of the radiation source respectively, and setting the maximum value to 1. The normalization formula is: Where Y is the normalized result and X is the data to be normalized.

3. The radiation source individual identification method based on deep sub-domain adaptation according to claim 2 is characterized in that: The conversion of the one-dimensional point data into an IQ waveform diagram includes: taking m data points as a data sample, changing the dimension of the one-dimensional IQ data according to the horizontal axis being time and the vertical axis being the size of the sample value, and obtaining a two-dimensional waveform diagram of the IQ two-channel data.

4. The method for radiation source individual identification based on deep sub-domain adaptation according to claim 3, characterized in that: The I-channel image and the corresponding Q-channel image are stitched together to form an I / Q stitching image, which includes: The I-channel and Q-channel two-dimensional waveforms are spliced ​​together, with the I-channel image on the left and the Q-channel image on the right.

5. The sub-domain adaptive radiation source individual identification method according to claim 2, characterized in that: Considering the additive white Gaussian noise channel scenario for the IQ mosaic image, the IQ mosaic images with different noise distributions are formed, including: The signal-to-noise ratio range is 0~kdB, and there are a total of k / 2 IQ splicing datasets with different noise environments at intervals of 2dB. The dataset with a signal-to-noise ratio of n dB is taken as the source domain data, and the remaining k / 2-1 noise datasets are taken as the target domain datasets.

6. The method for radiation source individual identification based on deep sub-domain adaptation according to claim 1, characterized in that: The local maximum mean difference adaptation layer is an improvement based on the maximum mean difference that measures the overall difference in the distribution between the source domain and the target domain data. By introducing the local maximum mean difference to perform conditional distribution difference calculation, it is minimized during the iteration process of the deep network, thereby narrowing the distribution differences of related subdomains within the same category.

7. The method for radiation source individual identification based on deep sub-domain adaptation according to claim 6, characterized in that: The local maximum mean difference expression is: in, and They are and The weight belonging to category c is calculated as: Where: y ic is the input vector Y i Category c label.

8. The method for radiation source individual identification based on deep sub-domain adaptation according to claim 5, characterized in that: For samples in the source domain, use the true label Calculate the weight of each sample as one-hot For unlabeled target domains, deep neural networks The output is a probability distribution, using the network to predict the label As the pseudo label of the target domain to calculate the weight of the target domain To achieve self-adaptation of high-level network layers, we need to know the activation value z l ; Given n s A labeled source domain D s , n t There is an unlabeled target domain D t , respectively obey the probability p and q, the deep neural network will produce in the l layer and The activation value of ; the sub-domain adaptive function is defined as: Where Z l is the feature extracted from the lth layer, l∈L={1,2,…,|L|}; During the model training process, LMMD in the above formula is used for domain adaptation loss, so that the objective function of network optimization is: Where J(·,·) is the cross entropy loss function; is the sub-domain adaptive function; Le is the total number of adaptive layers.

9. The method for radiation source individual identification based on deep sub-domain adaptation according to claim 5, characterized in that: The sub-domain adaptive model is trained using the I / Q splicing sample data of the source domain radiator signal with a signal-to-noise ratio of n. The I / Q splicing graphs of the radiators with k / 2-1 signal-to-noise ratios to be identified are input into the trained sub-domain adaptive model to perform transfer recognition on the target domain radiator individuals.

Images