Communication emitter identification method based on deep learning feature matching

Abstract

The application provides a communication radiation source identification method based on deep learning feature matching, obtains sampling data of a communication target radiation source electromagnetic signal, performs data cleaning and constructs a training data set and a test data set; the training data set is subjected to data enhancement; time-frequency conversion and normalization processing are performed on the training sample; the pretreated training sample is taken as model input, a feature extraction network of a radiation source identification model is constructed through two-dimensional convolution, one-dimensional convolution and hollow convolution; a target identification network of the radiation source identification model is constructed; through a back propagation algorithm, a cross-entropy and feature matching matrix correlation cost function are taken as targets for minimization, and the radiation source identification model is optimized; after pretreatment, the test sample of the test data set is input into the radiation source identification model for target identification test. In this way, the application can realize intelligent extraction of a communication radiation source target signal electromagnetic fingerprint feature and accurate identification of a radiation source target.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic spectrum monitoring technology, and in particular to a method for identifying communication radiation sources based on deep learning feature matching. Background Technology

[0002] Communication radiation source identification is a technology that matches collected and intercepted electromagnetic signals with the target individual itself. It is defined as "the ability to correlate the electromagnetic signal characteristics of a radiation source with the individual radiation source." Based on this technology, electromagnetic signals can be associated with information such as the carrier platform, user identity, and target parameters, thereby forming high-value electromagnetic situational information. Different communication radiation sources have variations in the production and assembly of their components. These hardware differences give even radiation sources of the same model and batch an inherent attribute that distinguishes them from other individuals, also known as "electromagnetic fingerprint characteristics." This attribute affects the radio signal but does not affect the radiation source's original communication or detection functions. Communication radiation source identification, through the selection of radiation source signal characteristics and high-precision measurement of characteristic parameters, discovers the unique features exhibited by the target individual in the signal, thereby achieving the purpose of individual identification of the radiation source target.

[0003] A typical individual target recognition process includes four modules: electromagnetic signal reception, signal preprocessing, feature extraction, and target recognition. Feature extraction aims to filter out redundant information in the signal data and provide more differentiated and representative individual characteristics of the radiation source. Existing methods generally require experts to conduct in-depth analysis of the differences in radiation source signal representations and are specifically designed accordingly. These methods can be mainly divided into transient feature extraction and steady-state feature extraction. Target recognition, on the other hand, constructs a radiation source classification and recognition model with strong classification and generalization capabilities based on existing data, feature extraction results, and prior knowledge. The main methods include traditional algorithms such as feature parameter matching and artificial intelligence-related algorithms such as clustering, statistical learning, and neural networks. However, with the significant improvement in electronic component manufacturing processes, the subtle differences among individual radiation sources have led to problems with traditional methods, such as fixed recognition targets, reliance on expert knowledge for feature extraction, lack of universality and scalability of algorithms, and insufficient recognition accuracy.

[0004] With the rapid development of artificial intelligence technology, its advantages in data representation, feature extraction, and classification prediction, combined with the characteristics of electromagnetic signal subtle feature extraction and target recognition tasks, provide new ideas and directions for intelligent and efficient communication radiation source identification technology. Therefore, research on intelligent identification methods for communication radiation sources based on deep learning is of great significance to the development of the electromagnetic spectrum monitoring field. Summary of the Invention

[0005] The purpose of this invention is to provide a communication radiation source identification method based on deep learning feature matching, which can realize intelligent extraction of electromagnetic fingerprint features of communication radiation source target signals and accurate identification of radiation source targets.

[0006] To achieve the above objectives, this invention provides a method for identifying communication radiation sources based on deep learning feature matching, comprising the following steps:

[0007] Acquire sampling data of electromagnetic signals from the radiation source of the communication target, perform data cleaning on the sampling data, and construct training datasets and test datasets;

[0008] Perform data augmentation on the training dataset;

[0009] The training samples in the training dataset are subjected to time-frequency transformation and normalization.

[0010] The preprocessed training samples are used as model input, and a feature extraction network for the radiation source identification model is constructed through two-dimensional convolution, one-dimensional convolution, and dilated convolution.

[0011] The target recognition network of the radiation source recognition model is constructed by using a feature matching matrix;

[0012] The radiation source identification model is optimized by using the backpropagation algorithm to minimize the cross-entropy and feature matching matrix correlation cost function.

[0013] After preprocessing, the test samples of the test dataset are input into the radiation source identification model for target identification testing.

[0014] Optionally, the step of acquiring sampling data of the electromagnetic signal of the communication target radiation source, cleaning the sampling data, and constructing training and testing datasets includes:

[0015] The electromagnetic signal of the radiation source of the communication target is sampled by the data acquisition equipment to obtain the sampling data of both I and Q channels;

[0016] Based on the signal quality parameters of the sampled data, the sampled data is cleaned, and training and testing datasets are constructed.

[0017] Optionally, the number of samples in the training dataset accounts for 70% of the total number of samples in the sampling data, and the number of samples in the test dataset accounts for 30% of the total number of samples in the sampling data, and the data in the training dataset and the test dataset are randomly shuffled.

[0018] Optionally, the data augmentation of the training dataset includes:

[0019] The training samples in the training dataset are subjected to signal-to-noise ratio enhancement and / or frequency enhancement to increase the number of training samples in the training dataset.

[0020] Optionally, the step of performing time-frequency transformation and normalization on the training samples in the training dataset includes:

[0021] The training samples in the training dataset are subjected to time-frequency transformation based on the following formula:

[0022] F=∫x(t)ω(t-τ)e -jωt dt; where x is the training sample and F is the frequency domain signal of the training sample after transformation;

[0023] The mean and variance of the training samples in the training dataset are normalized using the following formula:

[0024] Where μ is the mean of the training samples, and σ is the standard deviation of the training samples;

[0025] The time-domain and frequency-domain signals of the preprocessed training samples are fused to form the input data for the radiation source identification model.

[0026] Optionally, the feature extraction network includes convolutional layers and pooling layers. The convolutional layers use two-dimensional convolution, one-dimensional convolution, and dilated convolution to extract features from the preprocessed training samples. The pooling layers are used to reduce the dimension of the feature map.

[0027] The formula for calculating the two-dimensional convolution is:

[0028]

[0029] Among them, f conv_2D It is a two-dimensional convolution function, where the model input X, convolution kernel W, and output O are all two-dimensional matrices;

[0030] The formula for calculating the one-dimensional convolution is:

[0031]

[0032] Among them, f conv_1D For a one-dimensional convolution function, the model input X, the convolution kernel W, and the output O are all one-dimensional matrices;

[0033] The formula for calculating the dilated convolution is:

[0034]

[0035] Among them, f conv_dilatedHere, is the dilated convolution function, where di represents the dilation rate of the model input X in the first dimension, and dj represents the dilation rate of the model input X in the second dimension.

[0036] Optionally, the target recognition network for constructing the radiation source recognition model using a feature matching matrix includes:

[0037] The target recognition network of the radiation source identification model is constructed using a feature matching matrix and a softmax classifier; wherein, the feature matching matrix is ​​calculated based on the following formula:

[0038] yn=[ω1·on,ω2·on,...,ωk·on]

[0039] Where, ω k This represents the electromagnetic fingerprint feature corresponding to the k-th type of radiation source target in the feature matching matrix, o n Let y represent the feature vector obtained after the nth sample is processed by the feature extraction model. n The output vector for feature matching;

[0040] The prediction formula for the softmax classifier is as follows:

[0041]

[0042] Where, p n Let p be the radiation source target type prediction vector, K be the number of radiation source target types in the training dataset, and p be the radiation source target type prediction vector. nk The maximum value corresponds to the target type prediction value of the radiation source target type.

[0043] Optionally, optimizing the radiation source identification model using a backpropagation algorithm, with the objective of minimizing the cross-entropy and feature matching matrix correlation cost function, includes:

[0044] Using cross-entropy and feature matching matrix correlation as cost functions and minimizing the cost function as the objective, the radiation source identification model is trained using the Adam backpropagation algorithm until the radiation source identification model converges, so as to obtain the model parameters.

[0045] The cross-entropy and feature matching matrix correlation cost function are calculated based on the following formulas:

[0046]

[0047] Among them, L cross Let L be the cross-entropy cost function. cor The correlation cost function for the feature matching matrix. p is the true type vector of the input data. nK is the radiation source target type prediction vector; K is the number of radiation source target types, ω i The electromagnetic fingerprint feature is the feature matching matrix corresponding to the i-th type of radiation source target.

[0048] Optionally, the step of preprocessing the test samples of the test dataset and inputting them into the radiation source identification model for target identification testing includes:

[0049] The test samples in the test dataset are preprocessed using time-frequency transformation and normalization.

[0050] The preprocessed test sample is input into the radiation source identification model to obtain the corresponding output radiation source target type.

[0051] Optionally, the step of inputting the preprocessed test sample into the radiation source identification model to obtain the corresponding output radiation source target type further includes:

[0052] Calculate the model recognition accuracy based on the target recognition results output by the radiation source recognition model;

[0053] The model's recognition accuracy is calculated based on the following formula:

[0054]

[0055] Where, N correct N represents the number of correctly identified samples in the test dataset. sum This represents the total number of samples to be identified in the test dataset.

[0056] The communication radiation source identification method based on deep learning feature matching described in this invention first constructs training and test datasets based on collected data. Data augmentation and preprocessing are applied to the training dataset. Then, a feature extraction network and a target recognition network are constructed sequentially. The recognition model is trained using the training dataset, and the optimal recognition model parameters are obtained through continuous parameter tuning, thereby achieving target recognition on the test dataset. This invention leverages the advantages of deep learning in data processing and structured data feature extraction, and combines the time-frequency domain features of electromagnetic signals with a deep learning model to achieve intelligent extraction of electromagnetic fingerprint features of communication radiation source target signals and accurate identification of radiation source targets. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the steps of the communication radiation source identification method based on deep learning feature matching according to an embodiment of the present invention.

[0058] Figure 2This is a schematic diagram illustrating the specific process of the communication radiation source identification method based on deep learning feature matching provided in an embodiment of the present invention.

[0059] Figure 3 A schematic diagram of the convolutional structure used in the deep learning feature matching-based communication radiation source identification method provided in an embodiment of the present invention;

[0060] Figure 4 This is a diagram illustrating the architecture of the radiation source identification model trained by the deep learning feature matching-based communication radiation source identification method according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0062] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics, but not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0063] Furthermore, certain terms are used in the specification and subsequent claims to refer to specific components or parts. Those skilled in the art will understand that manufacturers may use different names or terms to refer to the same component or part. This specification and subsequent claims do not distinguish components or parts by differences in name, but rather by differences in function. The terms "comprising" and "including" used throughout the specification and subsequent claims are open-ended and should be interpreted as "including but not limited to." Additionally, the term "connection" here includes any direct and indirect electrical connection means. Indirect electrical connection means include connections made through other means.

[0064] Figure 1 This invention illustrates a method for identifying communication radiation sources based on deep learning feature matching, specifically applied to the identification of communication radiation source targets. The method includes the following steps:

[0065] S101: Acquire sampling data of electromagnetic signals from the radiation source of the communication target, clean the sampling data, and construct training and testing datasets. The electromagnetic signals from the radiation source of the communication target refer to the electromagnetic signals generated by the communication target. The communication target can be any object with communication capabilities, including fixed and mobile targets; furthermore, the communication target can include various targets in the air, on the ground, on the water surface, or underwater, such as aircraft, submersibles, cars, ships, etc.

[0066] The sampling data can be collected from electromagnetic signals generated by the actual communication target during use, or it can be collected by simulating electromagnetic signals.

[0067] In an optional implementation, step S101 includes: sampling the electromagnetic signal of the communication target radiation source using a data acquisition device to obtain sampling data for both I and Q channels; cleaning the sampling data according to the signal quality parameters of the sampling data, and constructing a training dataset and a test dataset. The data acquisition device is preferably a USRP software-defined radio device, but other types of radio receiving devices can also be used; this embodiment does not limit this. The sampling data is an electromagnetic signal, and the signal quality parameter of the sampling data is the signal-to-noise ratio (SNR) parameter. Specifically, the USRP software-defined radio device is used to acquire data from the communication radiation source target to obtain sampling data for both I and Q channels. To ensure the quality of the sampling data, a data cleaning method is further employed to remove data of poor quality based on the SNR of the electromagnetic signal.

[0068] In this embodiment, the training dataset comprises 70% of the total sampled data, and the test dataset comprises 30% of the total sampled data. The data in both the training and test datasets are randomly shuffled. Specifically, the training dataset is constructed using 70% of the cleaned total sample data, and the test dataset is constructed using the remaining 30%. The data in both datasets are then randomly shuffled. In this embodiment, the sample data dimension is [2, 1024].

[0069] Of course, in other examples, the number of samples between the training and test datasets can be adjusted depending on the specific application.

[0070] S102: Perform data augmentation on the training dataset. The purpose of data augmentation in this embodiment is to increase the sample size and diversity in the training dataset.

[0071] In an optional implementation, step S102 includes: performing signal-to-noise ratio (SNR) enhancement and / or frequency enhancement on the training samples in the training dataset to increase the number of training samples in the training dataset. The SNR enhancement involves randomly adding Gaussian white noise to the samples in the training dataset to increase the number of training samples; the frequency enhancement involves using methods such as undersampling or oversampling to increase the number of training samples in the training dataset.

[0072] S103: Perform time-frequency transformation and normalization on the training samples in the training dataset.

[0073] In specific implementation, step S103 of this embodiment specifically includes:

[0074] The time-frequency transformation of the training samples in the training dataset is performed based on the following formula:

[0075] F=∫x(t)ω(t-τ)e -jωt dt; where x is the training sample and F is the frequency domain signal of the training sample after transformation;

[0076] The mean and variance of the training samples in the training dataset are normalized using the following formula:

[0077] Where μ is the mean of the training samples and σ is the standard deviation of the training samples;

[0078] The time-domain and frequency-domain signals of the preprocessed training samples are fused to form the input data for the radiation source identification model.

[0079] S104: Using the preprocessed training samples as model input, a feature extraction network for the radiation source identification model is constructed through two-dimensional convolution, one-dimensional convolution, and dilated convolution. Step S104 uses two-dimensional convolution, one-dimensional convolution, and dilated convolution specifically to achieve multi-dimensional feature extraction.

[0080] In specific implementation, the feature extraction network includes convolutional layers and pooling layers. The convolutional layers use two-dimensional convolution, one-dimensional convolution, and dilated convolution to extract features from the preprocessed training samples, such as... Figure 3 As shown; pooling layers are used to reduce the dimension of feature maps;

[0081] The formula for calculating the two-dimensional convolution is:

[0082]

[0083] Among them, f conv_2D It is a two-dimensional convolution function, where the model input X, convolution kernel W, and output O are all two-dimensional matrices;

[0084] The formula for calculating the one-dimensional convolution is:

[0085]

[0086] Among them, f conv_1D For a one-dimensional convolution function, the model input X, the convolution kernel W, and the output O are all one-dimensional matrices;

[0087] The formula for calculating the dilated convolution is:

[0088]

[0089] Among them, f conv_dilated Here, is the dilated convolution function, where di represents the dilation rate of the model input X in the first dimension, and dj represents the dilation rate of the model input X in the second dimension.

[0090] The pooling layer selects the maximum value from the corresponding regions of the convolutional feature map using a pooling window of size w*h, where w is the row of the matrix and h is the column of the matrix.

[0091] In this embodiment, a feature extraction network based on four convolutional layers is constructed. The first convolutional layer uses a combination of two-dimensional convolution and dilated convolution. The kernel sizes of the two-dimensional convolutions are [2, 3] and [2, 7], with a total of 128 kernels. The kernel size of the dilated convolutions is [2, 3], with a total of 128 kernels and a dilation rate of [0, 2]. The second convolutional layer uses a combination of one-dimensional convolution and dilated convolution. The kernel sizes of the two-dimensional convolutions are [1, 3], with a total of 128 kernels. The kernel size of the dilated convolutions is [1, 3], with a total of 128 kernels and a dilation rate of [0, 1]. The remaining convolutional layers all have kernel sizes of [1, 3], with a total of 64 kernels.

[0092] S105: Construct the target recognition network of the radiation source recognition model using the feature matching matrix.

[0093] In specific implementation, step S105 includes: constructing the target recognition network of the radiation source recognition model using a feature matching matrix and a softmax classifier; wherein, the feature matching matrix is ​​calculated based on the following formula:

[0094] yn=[ω1·on,ω2·on,...,ωk·on]

[0095] Where, ω k This represents the electromagnetic fingerprint feature corresponding to the k-th type of radiation source target in the feature matching matrix, o n Let y represent the feature vector obtained after the nth sample is processed by the feature extraction model. n The output vector for feature matching;

[0096] The prediction formula for the softmax classifier is as follows:

[0097]

[0098] Where, p n Let p be the radiation source target type prediction vector, K be the number of radiation source target types in the training dataset, and p be the radiation source target type prediction vector. nk The maximum value corresponds to the target type prediction value of the radiation source target type.

[0099] The target recognition network uses a softmax classifier to identify different types of targets. The softmax classifier is used to classify the output value y of the k-th class. nk Convert to a probability distribution with a range of [0, 1] and a sum of 1, p nk The target type corresponding to the maximum value is the predicted value of the radiation source target type.

[0100] In this embodiment, the electromagnetic fingerprint feature vector ω corresponding to the k-th type of radiation source target in the feature matching matrix is... k The dimension is 64, and the number of nodes in the softmax layer is 36.

[0101] S106: Optimize the radiation source identification model by using the backpropagation algorithm, with the goal of minimizing the cross-entropy and feature matching matrix correlation cost function.

[0102] In specific implementation, step S106: using cross-entropy and feature matching matrix correlation as cost functions and minimizing the cost function as the objective, the radiation source identification model is trained using the Adam backpropagation algorithm until the radiation source identification model converges, so as to obtain the model parameters.

[0103] The cross-entropy and feature matching matrix correlation cost function are calculated based on the following formulas:

[0104]

[0105] Among them, L cross Let L be the cross-entropy cost function. cor The correlation cost function for the feature matching matrix. p is the true type vector of the input data. n K is the radiation source target type prediction vector; K is the number of radiation source target types, ω i This refers to the electromagnetic fingerprint feature corresponding to the i-th type of radiation source target in the feature matching matrix. In this embodiment, the number of radiation sources is K = 36.

[0106] This embodiment aims to minimize the cost function and uses the Adam backpropagation algorithm to train the radiation source identification model until the identification model converges, thus obtaining the model parameters; preferably, the initial learning rate is 0.001.

[0107] S107: After preprocessing the test samples of the test dataset, input them into the radiation source identification model for target identification testing. This step is to test the target identification accuracy of the radiation source identification model so that the final model with a target identification accuracy meeting the preset requirements can be applied to target identification of the electromagnetic signals of radiation sources.

[0108] In specific implementation, step S107 includes: performing time-frequency transformation and normalization preprocessing on the test samples in the test dataset; inputting the preprocessed test samples into the radiation source identification model to obtain the corresponding output radiation source target type. That is, the test samples in the test dataset are preprocessed in the same way as the training samples, and the preprocessed test samples are input into the trained radiation source identification model so that the radiation source identification model outputs the corresponding target identification result, i.e., the radiation source target type.

[0109] Furthermore, step S107 also includes: calculating the model recognition accuracy based on the target recognition result output by the radiation source recognition model;

[0110] The model's recognition accuracy is calculated based on the following formula:

[0111]

[0112] Where, N correct N represents the number of correctly identified samples in the test dataset. sum This represents the total number of samples to be identified in the test dataset.

[0113] See Figure 2 and Figure 4 The specific process of the communication radiation source identification method based on deep learning feature matching provided in this embodiment is as follows:

[0114] 1. Obtain sample data through data collection and construct training and testing datasets respectively;

[0115] 2. Perform data augmentation on the training dataset to increase sample size and diversity;

[0116] 3. Perform data preprocessing on the training dataset so that the time-domain signal and frequency-domain signal of the preprocessed sample data are fused to form the input data of the radiation source identification model;

[0117] 4. Construct a feature extraction network based on the preprocessed training samples;

[0118] 5. Construct an object recognition network based on the preprocessed training samples;

[0119] 6. Optimize the initial model built based on the feature extraction network and the target recognition network to obtain the radiation source recognition model;

[0120] 7. After preprocessing, the test samples in the test dataset are used as input to the radiation source identification model to obtain the target identification results and calculate the target identification accuracy.

[0121] In summary, the communication radiation source identification method based on deep learning feature matching described in this invention leverages the advantages of deep learning in feature extraction by combining electromagnetic fingerprint features with a deep learning model. This enables intelligent extraction of subtle features of electromagnetic signals from radiation sources, eliminating the limitations of traditional methods that rely on prior knowledge. Furthermore, the use of data augmentation methods enhances the representational capabilities of training data, effectively improving the robustness and accuracy of the communication radiation source identification model.

[0122] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0123] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

[0127] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A method for identifying communication radiation sources based on deep learning feature matching, characterized in that, Including the following steps: Acquire sampling data of electromagnetic signals from the radiation source of the communication target, perform data cleaning on the sampling data, and construct training datasets and test datasets; Perform data augmentation on the training dataset; The training samples in the training dataset are subjected to time-frequency transformation and normalization. The preprocessed training samples are used as model input, and a feature extraction network for the radiation source identification model is constructed through two-dimensional convolution, one-dimensional convolution, and dilated convolution. The target recognition network of the radiation source recognition model is constructed by using a feature matching matrix; Using feature matching matrix and softmax A classifier is used to construct the target recognition network of the radiation source identification model; wherein, the feature matching matrix is ​​calculated based on the following formula: ； in, Indicates the feature matching matrix corresponding to the first k Electromagnetic fingerprint characteristics corresponding to radiation source-like targets. Indicates the first n The feature vector obtained after processing each sample by the feature extraction model The output vector for feature matching; The softmax The prediction formula for the classifier is as follows: ； in, This is the prediction vector for the radiation source target type. The number of radiation source target types in the training dataset. The target type corresponding to the maximum value is the predicted value of the radiation source target type; The radiation source identification model is optimized using the backpropagation algorithm, with the objective of minimizing the cross-entropy and feature matching matrix correlation cost function, including: Using cross-entropy and feature matching matrix correlation as cost functions and minimizing the cost function as the objective, the radiation source identification model is trained using the Adam backpropagation algorithm until the radiation source identification model converges, so as to obtain the model parameters. The cross-entropy and feature matching matrix correlation cost function are calculated based on the following formulas: ； in, Let cross-entropy cost function be used. The correlation cost function for the feature matching matrix. The vector representing the true type of the input data; This is a prediction vector for the target type of the radiation source; The number of radiation source target types. For the feature matching matrix of the th Electromagnetic fingerprint characteristics corresponding to radiation source-like targets; After preprocessing, the test samples of the test dataset are input into the radiation source identification model for target identification testing.

2. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The process of acquiring sampling data of electromagnetic signals from the radiation source of the communication target, cleaning the sampling data, and constructing training and testing datasets includes: The electromagnetic signal of the radiation source of the communication target is sampled by the data acquisition equipment to obtain the sampling data of both I and Q channels; Based on the signal quality parameters of the sampled data, the sampled data is cleaned, and training and testing datasets are constructed.

3. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The training dataset contains 70% of the total number of samples in the sampled data, and the test dataset contains 30% of the total number of samples in the sampled data. The data in the training dataset and the test dataset are randomly shuffled.

4. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The data augmentation of the training dataset includes: The training samples in the training dataset are subjected to signal-to-noise ratio enhancement and / or frequency enhancement to increase the number of training samples in the training dataset.

5. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The step of performing time-frequency transformation and normalization on the training samples in the training dataset includes: The training samples in the training dataset are subjected to time-frequency transformation based on the following formula: ;in, x The training samples are as follows. F The frequency domain signal after the training sample is transformed; The mean and variance of the training samples in the training dataset are normalized using the following formula: ;in, μ The mean of the training samples is denoted as . The standard deviation of the training samples; The time-domain and frequency-domain signals of the preprocessed training samples are fused to form the input data for the radiation source identification model.

6. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The feature extraction network includes convolutional layers and pooling layers. The convolutional layers use two-dimensional convolution, one-dimensional convolution, and dilated convolution to extract features from the preprocessed training samples. The pooling layers are used to reduce the dimension of the feature maps. The formula for calculating the two-dimensional convolution is: ； in, For a two-dimensional convolution function, the model input is... X convolution kernel W and output O All are two-dimensional matrices; The formula for calculating the one-dimensional convolution is: ； in, For a one-dimensional convolution function, the model input is... X convolution kernel W and output O All are one-dimensional matrices; The formula for calculating the dilated convolution is: ； in, It is a dilated convolution function. di Represents model input X In the first dimension, the void ratio DJ Represents model input X The void ratio in the second dimension.

7. The communication radiation source identification method based on deep learning feature matching according to claim 1, characterized in that, The step of preprocessing the test samples of the test dataset and inputting them into the radiation source identification model for target identification testing includes: The test samples in the test dataset are preprocessed using time-frequency transformation and normalization. The preprocessed test sample is input into the radiation source identification model to obtain the corresponding output radiation source target type.

8. The communication radiation source identification method based on deep learning feature matching according to claim 7, characterized in that, The step of inputting the preprocessed test sample into the radiation source identification model to obtain the corresponding output radiation source target type further includes: Calculate the model recognition accuracy based on the target recognition results output by the radiation source recognition model; The model's recognition accuracy is calculated based on the following formula: ； in, The number of correctly identified samples in the test dataset. This represents the total number of samples to be identified in the test dataset.

Images