Unsupervised Domain Adaptation Radiation Source Individual Recognition Method Based on Independence Criterion

Abstract

The present invention discloses an unsupervised domain adaptation radiation source individual recognition method based on the independence criterion. First, an identification model is constructed, which includes a feature extractor, a source domain classifier, and a target domain classifier, and the parameters of the source domain classifier and the target domain classifier are shared; then the identification model is trained, and the total loss function used during training includes a classification loss function and a distribution distance metric function for cross-domain samples constructed based on the Hilbert-Schmidt independence criterion; finally, the target domain data to be recognized is input into the feature extractor and the target domain classifier to obtain a classification result. The present invention adopts unsupervised domain adaptation technology, which is applicable to the situation where the target domain only contains unlabeled data, thus significantly reducing the dependence on high-quality labeled data in the target domain, and at the same time having the advantages of high accuracy and good robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of radiation source identification, and particularly to a method for identifying individual radiation sources. Background Art

[0002] Specific Emitter Identification (SEI) technology accurately identifies individual radiation sources that emit these signals by extracting unique radio frequency fingerprint features from captured electromagnetic signals and utilizing existing knowledge. Radio frequency fingerprints are unique identifiers formed by inherent variations during the hardware manufacturing process, so they are inherently unavoidable and difficult to replicate. The extraction and classification of radio frequency fingerprints are the core of specific emitter identification technology.

[0003] Currently, SEI technology is mainly divided into two types: one is the traditional radiation source identification technology that requires manual intervention; the other is the automatic radiation source identification technology based on deep neural networks.

[0004] Traditional radiation source identification technology starts from the signal generation mechanism and involves features in multiple aspects such as transform domain information, modulation domain information, and signal image information through manually designed feature extraction methods. However, this method is easily affected by human factors and has limitations in its ability to characterize individual radiation source information.

[0005] In recent years, the wide application of deep neural networks in multiple fields has demonstrated its powerful non - linear feature extraction ability, bringing innovative feature extraction methods to the field of specific radiation source identification. At the same time, the rapid development of the Internet of Things technology has led to a sharp increase in the types and quantities of devices, posing new challenges to traditional data - processing methods and making it increasingly difficult for them to meet the extensive data - processing requirements. Against this background, the continuous progress of artificial intelligence technology is driving the transformation of specific emitter identification technology from traditional methods to more efficient deep - learning methods. Deep - learning technology can effectively mine deep - level radio frequency features, achieve fast and efficient feature extraction, and significantly improve the accuracy of individual radiation source identification. Nevertheless, deep neural networks still face the problem of scarce high - quality labeled data in practical applications, which to a certain extent limits their wide application in the field of radiation source identification. Summary of the Invention

[0006] The present invention proposes an unsupervised domain - adaptation radiation source individual identification method based on the independence criterion, and its purpose is to solve the problem of scarce high - quality labeled data existing in the practical application of deep neural networks.

[0007] The technical solution of the present invention is as follows:

[0008] An unsupervised domain - adaptation radiation source individual identification method based on the independence criterion, the steps include:

[0009] Step 1: Construct an identification model;

[0010] The identification model includes a feature extractor, a source domain classifier, and a target domain classifier; the feature extractor is used to extract first features from source domain data and also to extract second features from target domain data; the source domain classifier is used to obtain a source domain classification result based on the first features, and the target domain classifier is used to obtain a target domain classification result based on the second features; the source domain classifier and the target domain classifier share parameters;

[0011] Step 2: Train the identification model; the total loss function used during training includes a classification loss function and a distribution distance metric function for cross-domain samples constructed based on the Hilbert - Schmidt Independence Criterion;

[0012] After training is completed, input the target domain data to be identified into the feature extractor to obtain the corresponding second features, and then input the second features into the target domain classifier to obtain a classification result.

[0013] As a further improvement to the above-mentioned unsupervised domain adaptation radiation source individual identification method based on the independence criterion: the feature extractor includes four residual layers, and each residual layer includes four residual blocks; each residual block includes three groups of alternately arranged convolutional layers and BN layers, and also includes an activation function layer at the end.

[0014] As a further improvement to the above-mentioned unsupervised domain adaptation radiation source individual identification method based on the independence criterion: both the source domain classifier and the target classifier include three fully connected layers and one Softmax layer.

[0015] As a further improvement to the above-mentioned unsupervised domain adaptation radiation source individual identification method based on the independence criterion, the distribution distance metric function is:

[0016] ;

[0017] where is a hyperparameter, is the number of sample pairs in the current batch, and the sample pair includes one source sample and one target domain sample, represents taking the trace of a matrix, and are the kernel matrices calculated based on the source domain sample set in the current batch and the target domain sample set in the current batch respectively;

[0018] The matrix is:

[0019] ;

[0020] Among them, is the identity matrix, is a column vector with all elements being 1, .

[0021] As a further improvement of the above-mentioned unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, the calculation method of the kernel matrix is as follows: Calculate the kernel function values of any two samples and in the sample set:

[0022] ;

[0023] Among them, is the scale parameter;

[0024] Then take as the element value of the th row and the th column in the kernel matrix.

[0025] As a further improvement of the above-mentioned unsupervised domain adaptation radiation source individual recognition method based on the independence criterion: Dynamically adjust the scale parameter, and determine the variance of all samples in the current sample set as the scale parameter when calculating the kernel function.

[0026] As a further improvement of the above-mentioned unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, the classification loss function is:

[0027] ;

[0028] Among them, is the number of source domain samples in the current batch, is the number of categories; is an indication label indicating that the th source domain sample belongs to the th category. It is 1 if it belongs to the th category and 0 if it does not belong to the th category; is the probability value that the th source domain sample predicted by the source domain classifier belongs to the th category.

[0029] As a further improvement of the above-mentioned unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, the total loss function used during training in step 2 is:

[0030] ;

[0031] Among them, is the classification loss function, is the distribution distance metric function, is a hyperparameter.

[0032] As a further improvement to the unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, after each batch of training data is input in step 2, the total loss function is calculated once and the parameters of the recognition model are iterated. Each batch contains sample pairs, and each sample pair contains a target domain sample, a source domain sample, and the corresponding label data.

[0033] As a further improvement to the unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, the target domain samples and source domain samples are constructed by adding Gaussian white noise with different signal-to-noise ratios to the originally collected samples.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The present invention adopts the unsupervised domain adaptation technology, which is applicable to the situation where the target domain only contains unlabeled data, thus significantly reducing the dependence on high-quality labeled data in the target domain.

[0036] 2. The present invention uses the independence criterion (HSIC) to measure the similarity between the source domain and the target domain, and proposes a cross-domain similarity metric function in the form of a negative exponent, thereby improving the ability to extract similarity features, and further improving the accuracy of radiation source individual recognition.

[0037] 3. The present invention introduces a parameter dynamic adjustment mechanism in the cross-domain similarity metric, improving the accuracy of cross-domain similarity measurement for different samples and enhancing the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is the overall architecture diagram of the recognition model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The technical solution of the present invention will be described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0040] An unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, the steps include:

[0041] Step 1: Construct a recognition model.

[0042] As Figure 1 , the recognition model includes a feature extractor, a source domain classifier, and a target domain classifier.

[0043] The feature extractor is used to extract the first feature from the source domain data and also to extract the second feature from the target domain data.

[0044] The feature extractor is dedicated to learning a mapping to extract high-level features that can be used for classification from the input sample data. This process can be expressed as:

[0045] ;

[0046] Among them, represents the mapping function learned by the feature extractor, represents the sample data with a dimension of represents the variable parameter of the mapping function, that is, the parameter to be trained in the feature extractor, represents the -dimensional feature that has obvious inter-class classification and domain-invariant characteristics.

[0047] Specifically, the feature extractor includes four residual layers, and each residual layer contains four residual blocks. Each residual block includes three groups of alternately arranged convolutional layers and BN layers, and also includes an activation function layer at the end.

[0048] The source domain classifier is used to obtain the source domain classification result according to the first feature, and the target domain classifier is used to obtain the target domain classification result according to the second feature.

[0049] Specifically, both the source domain classifier and the target classifier include three fully connected layers and one Softmax layer. A Dropout layer is also added to the classifier network to prevent overfitting during training. The Dropout layer reduces the dependence between neurons by randomly setting the outputs of some neurons in the network to 0 during training.

[0050] The source domain classifier and the target domain classifier share parameters.

[0051] Step 2: Train the recognition model.

[0052] Before training, prepare the training samples in the following way:

[0053] (1) First, preprocess the collected raw data. In this embodiment, the collected data is 2.4G WiFi signals, the WiFi device model is ESP8266, set to the 802.11b WiFi standard, and the signal bandwidth is 20MHz. The preprocessing includes re-segmenting the data, normalizing the data, and integrating the data into a three-dimensional shape (20000, 2, 1000).

[0054] ​(2) The above dataset contains the signals of 10 individual radiation sources, which are respectively encoded as 0 - 9, and the corresponding label data is generated for the corresponding signal sample data, with a shape of (20000, 1).

[0055] (3) The processed data and the corresponding label data are put in one - to - one correspondence, and then randomly divided into source - domain and target - domain datasets according to a ratio of 1:1.

[0056] (4) Gaussian white noise with different signal - to - noise ratios (20dB, 18dB, 16dB, and 14dB) is added to the source - domain dataset respectively to simulate different data distributions. The shape of the source - domain data is (10000, 2, 1000), and the shape of the label data is (10000, 1).

[0057] (5) Gaussian white noise with a signal - to - noise ratio of 4dB is added to the original target - domain dataset. Then, the target - domain dataset is further divided into a training set and a test set at a ratio of 8:2. The label data of the target - domain training set is removed, while the label data of the test set (2000, 1) is retained. The training set is used for training the model, and the test set is used to evaluate the performance of the trained model on the target domain.

[0058] (6) The source - domain samples in the source - domain dataset and the target - domain samples in the target - domain training set are randomly combined into a number of sample pairs. Each sample pair contains a target - domain sample, a source - domain sample, and the corresponding label data. Then all sample pairs are divided into several batches, and each batch contains sample pairs.

[0059] During training, the total loss function is calculated once after each batch of training data is input, and the parameters of the recognition model are iterated. The total loss function includes a classification loss function and a distribution distance metric function constructed based on the Hilbert - Schmidt Independence Criterion, specifically:

[0060] ;

[0061] where, is the classification loss function, is the distribution distance metric function, is a hyperparameter.

[0062] The classification loss function is:

[0063] ;

[0064] where, is the number of source - domain samples in the current batch, is the number of classes; represents that the - th source - domain sample belongs to the An indication label for a category, where 1 indicates belonging to the category and 0 indicates not belonging to the category; The probability value that the th source domain sample predicted by the source domain classifier belongs to the category.

[0065] The distribution distance metric function is:

[0066] ;

[0067] where is a hyperparameter, is the number of sample pairs in the current batch, and the sample pair contains one source sample and one target domain sample, represents taking the trace of a matrix, and are the kernel matrices calculated based on the source domain sample set in the current batch and the target domain sample set in the current batch, respectively.

[0068] Matrix is:

[0069] ;

[0070] where is the identity matrix, is a column vector with all elements being 1, .

[0071] Furthermore, the calculation method of the kernel matrix is: calculate the kernel function value of any two samples and in the sample set:

[0072] ;

[0073] where is the scale parameter;

[0074] Then take as the element value of the th row and the th column in the kernel matrix.

[0075] The magnitude of the scale parameter determines the sensitivity of the Gaussian kernel's variation. In practical applications, if the scale parameter is chosen too small, any data will be linearly separable after being mapped to the Hilbert space, leading to overfitting. If the scale parameter is chosen too large, the data will be mapped to a low-dimensional space, resulting in underfitting. Therefore, for finite data, HSIC places more emphasis on the scale relationship. During the network training process, the sample data comes from two domains with different distributions and varies with each batch. Thus, the scale parameter needs to change dynamically. In the present invention, the variance of all samples in the current sample set is used as the scale parameter when calculating the kernel function to ensure that the most suitable scale parameter is selected for each batch of data, realizing the dynamic change of the scale parameter.

[0076] In this embodiment, the Adam function is set as the optimization function, the number of training epochs is set to 100, the transfer weight is set to 20, the random seed is 10, and the initial learning rate lr is set to 0.001. It is set to 10.

[0077] Step 3: After training is completed, the target domain data to be recognized is input into the feature extractor to obtain the corresponding second feature, and then the second feature is input into the target domain classifier to obtain the classification result.

[0078] Table 1 shows the recognition results of radiation source individuals using different methods for the above test set. Under the four transfer scenarios, the average recognition accuracy of this method is 84%, which is 41.77% higher than the baseline method and 5.77% higher than the sub-optimal MRAN (Multi-Representation Adaptation Network) method. Under each transfer scenario, the recognition accuracy of this method is above 83%. Compared with other transfer methods being compared, the transfer effect of this method is more stable. In the scenario of transferring from the source domain to the target domain with a signal-to-noise ratio of 18 dB, the performance of the baseline method drops the most. This shows that in this scenario, the cross-domain data samples differ greatly, but the similarity measurement method based on the independence criterion adopted by this method can still well mine the similarity between cross-domain samples, and can obtain general features with good intra-class aggregation and inter-class separability from the source domain to transfer knowledge from the source domain to the target domain.

[0079] Table 1 - Comparison of recognition accuracies between the method of the present invention and other methods

[0080]

[0081] Table 2 shows the ablation experiment results of the similarity measurement method. In this experiment, only the distribution distance measurement function in the proposed model was changed to explore the impact of different distribution distance measurement functions on the model performance. This experiment compared the impacts of three distance measurement functions, namely HSIC, MMD, and CORAL, on the model performance. The average recognition accuracy of the method (HSIC) was the highest, reaching 84%, followed by MMD at 74.90%, and CORAL was the worst at 74.46%. When the similarity between the source domain and target domain samples was small, the network performance of using MMD and CORAL for similarity measurement was higher than that of cross-domain samples with larger similarities, while the network performance of using HSIC changed little. This indicates that HSIC can measure the distance between different distributions well, effectively improve the performance of the target domain model, and reduce the dependence on the similarity of cross-domain samples themselves.

[0082] Table 2 - Impacts of Three Distance Measurement Functions on Model Performance

[0083]

[0084] It should be noted that for those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. An unsupervised domain adaptation radiation source individual recognition method based on the independence criterion, characterized in that the steps Including: Step 1: Construct an identification model; The identification model includes a feature extractor, a source domain classifier, and a target domain classifier; the feature extractor is used to extract first features from source domain data and also to extract second features from target domain data; the source domain classifier is used to obtain a source domain classification result based on the first features, and the target domain classifier is used to obtain a target domain classification result based on the second features; The source domain classifier and the target domain classifier share parameters; Step 2: Train the identification model; including: Prepare training samples, preprocess the collected 2.4G Wi-Fi signals, encode each radiation source individual and generate corresponding label data for the corresponding signal sample data, pair the processed data with the corresponding label data one by one, and then randomly divide them into a source domain data set and a target domain data set; The target domain samples and the source domain samples are constructed by adding Gaussian white noise with different signal-to-noise ratios to the originally collected samples; The total loss function used during training includes a classification loss function and a distribution distance metric function for cross-domain samples constructed based on the Hilbert-Schmidt independence criterion; The distribution distance metric function is: ； Among them, is a hyperparameter, is the number of sample pairs in the current batch, and the sample pair includes a source domain sample and a target domain sample, represents the trace of a matrix, and are respectively the kernel matrix calculated based on the source domain sample set in the current batch and the kernel matrix calculated based on the target domain sample set in the current batch; Matrix is as follows: ； Among them, is the identity matrix, is a column vector with all elements being 1, ; The calculation method of the kernel matrix is as follows: Calculate the kernel function values of any two samples in the sample set and : ； wherein, is a scale parameter; Then, is used as the element value of the -th row and the -th column in the kernel matrix; Dynamically adjust the scale parameter, and determine the variance of all samples in the current sample set as the scale parameter when calculating the kernel function; The total loss function used during training in Step 2 is: ； wherein, is the classification loss function, is the distribution distance metric function, is a hyperparameter; After training is completed, input the target domain data to be identified into the feature extractor to obtain the corresponding second features, and then input the second features into the target domain classifier to obtain a classification result for identifying the radiation source individuals.

2. The unsupervised domain adaptation radiation source individual recognition method based on the independence criterion according to claim 1, wherein: The feature extractor contains four residual layers, and each residual layer contains four residual blocks; each residual block includes three groups of alternately arranged convolutional layers and BN layers, and also includes an activation function layer at the end.

3. The unsupervised domain adaptation radiation source individual recognition method based on the independence criterion according to claim 1, characterized in that: Both the source domain classifier and the target classifier include three fully connected layers and one Softmax layer.

4. The unsupervised domain adaptation radiation source individual recognition method based on the independence criterion according to claim 1, characterized in that The classification loss function is: ； wherein, is the number of source domain samples in the current batch, is the number of categories; is an indication label indicating that the th source domain sample belongs to the th category. It is 1 if it belongs to the th category and 0 if it does not belong to the th category; is the probability value that the th source domain sample predicted by the source domain classifier belongs to the th category.

5. The unsupervised domain adaptation radiation source individual identification method based on the independence criterion according to any one of claims 1 to 4, characterized in that After each batch of training data is input in Step 2, the total loss function is calculated once and the parameters of the recognition model are iterated. Each batch contains sample pairs. Each sample pair contains a target domain sample, as well as a source domain sample and the corresponding label data.

Images