Wireless communication signal open set identification method based on convolution block attention and reciprocity point

Abstract

The invention relates to the technical field of wireless communication, and discloses a wireless communication signal open set identification method based on convolution block attention and reciprocity points, comprising: receiving wireless communication signal data including K types of known signal data and U types of unknown signal data; marking known signal data to form a training data set; forming an open test data set by the known signal data and the unknown signal data; converting signal data in the training data set and the open test data set into a time-frequency graph; constructing a reciprocity point model, and obtaining potential feature representation of an unknown signal sample in an out-of-class space corresponding to each piece of known class signal data; training a classification model based on the training data set; and inputting the open test data set into the classification model to output K + 1 type labels of the signal type corresponding to the maximum reciprocity point distance. According to the technical scheme, abnormal and unknown signals can be identified under the influence of noise and interference in a complex open electromagnetic environment.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a method for open set recognition of wireless communication signals based on convolutional block attention and reciprocity points. Background Technology

[0002] In complex open electromagnetic environments, it is necessary to classify and identify signals, understand their properties, and analyze the risks in open spaces.

[0003] For signal classification and recognition, deep learning achieves excellent results in identifying image data or categorical data. Traditional supervised signal recognition algorithms are based on limited prior knowledge of known classes. During testing, the learning model matches the input instance with the class with the highest posterior probability; however, it is impossible to use all existing classes as the training set. Therefore, the learning model treats all new class instances as known classes, leading to a degraded system performance. Unsupervised models, on the other hand, cannot identify and classify known classes.

[0004] In practical applications, open-set recognition is required: that is, classifying known data while simultaneously detecting data from unknown classes; during the recognition process, it is crucial to reduce the empirical classification risk of labeled known data and the open-space risk of potentially unknown data. Existing open-set recognition models typically use generative adversarial networks, autoencoders, or mainstream models to generate unknown or known samples to help the classifier learn the decision boundary between known and unknown samples. However, these methods do not consider the depth distribution of unknown classes in the learning model, leading to potential open-space risks. Prototypes learned only from known classes may converge in the space of unknown classes during training, making it impossible to distinguish between known and unknown classes. Therefore, training should not only model how to learn known classes but also the space of potential unknown classes.

[0005] Therefore, there is a need for an open-set identification method for wireless communication signals that can effectively improve the identification accuracy of known signal types and the detection accuracy of unknown signals in complex open electromagnetic environments. Summary of the Invention

[0006] To achieve the above objectives, this application provides a method for open set recognition of wireless communication signals based on convolutional block attention and reciprocity points, comprising the following steps: Receive wireless communication signal data, which includes K types of known signal data and U types of unknown signal data; label the K types of known signal data to form a training dataset. The open test dataset is constructed by combining K types of known signal data and U types of unknown signal data. Training dataset The feature space to which it belongs is Signal category The feature space of the dataset is used This indicates that the corresponding open space is Unknown signal data in the open test dataset are not labeled. training dataset and open test dataset The signal data in the diagram is converted into a time-frequency graph; Construct based on training dataset The reciprocity point model is used to obtain the latent feature representation of unknown signal samples in the out-of-class space corresponding to each known class of signal data; Based on training dataset A classification model is trained, which is based on an adversarial network model and a feature extraction network. Open test dataset The input is fed into the classification model, which outputs the K+U class labels of the signal class corresponding to the maximum reciprocity distance, where K is the number of known signal classes. "To identify categories that are classified as unknown."

[0007] Among them, the known signal data includes legitimate signals and malicious / illegal signals; legitimate signals are labeled as normal wireless communication signals; malicious / illegal signals are labeled as abnormal wireless communication signals.

[0008] Training dataset It contains training data with K known classes, represented as: ,in yes Tags; Open test dataset Represented as: , where k+1 to k+u represent unknown signals of class U.

[0009] Furthermore, the conversion to a time-frequency plot is achieved using smoothed pseudo-Wigner-Ville time-frequency analysis, as shown below: , in, It is a time difference variable. Represents a time variable. Represents frequency variables. Represents the signal function. Allows cross terms that oscillate parallel to the time axis to execute a time-smoothed real symmetric window function. Cross terms that oscillate parallel to the frequency axis are allowed to perform frequency-smoothed real symmetric window functions.

[0010] Furthermore, construct a system based on the training dataset. The reciprocal point model includes the following steps: Reciprocity points for wireless communication signal class K Modeling is performed where the reciprocity points of communication signal category K are... yes Out-of-class feature space Potential representation samples ; The distance between a given signal sample and a reciprocal point is defined using a trading point model. Based on the distance between a given signal sample and a reciprocal point, identification and classification learning is performed to determine the known and unknown spaces.

[0011] The distance between a given signal sample and a reciprocal point is determined by both the Euclidean distance and the cosine distance, expressed as: , in, This represents the distance between a given signal sample and a reciprocal point. This represents the feature extraction function, which is the function that maps signal samples to the feature space, where m is the dimension.

[0012] Furthermore, recognition and classification learning includes: Define the classification probability and normalize it using the softmax function, expressed as: ,in, It is a hyperparameter that controls the distance-probability transformation; Recognition and classification learning is achieved by minimizing the reciprocal point classification loss based on the known negative log probabilities of class K, expressed as: ,in These are the parameters of the feature extraction function.

[0013] Among them, based on the training dataset Training a classification model includes the following steps: will come from Samples and reciprocal points Distance constraints between scope; A classification model is defined, which includes an adversarial network model and a feature extraction network. The adversarial network model generates confusing samples as unknown signal data. A convolutional block attention module is introduced into the feature extraction network to make it pay more attention to the feature information of the energy focus areas in the time-frequency plot. Using training dataset Train the classification model.

[0014] The adversarial network model includes a discriminator D, a generator G, and a classifier C. The discriminator D represents the probability that a sample comes from the true distribution or the false distribution, and the generator G generates a sample. From the prior distribution Mapped to the output, the classifier C represents the probability that a sample belongs to each known class. Given a prior distribution... Generate samples and known signal samples The discriminator D is optimized to distinguish between real and generated fake samples.

[0015] Furthermore, the adversarial network model is generated through joint training, including the following steps: Mixed distribution samples are decoupled by auxiliary batch normalization; the auxiliary batch normalization refers to establishing independent batch normalization statistics for different domain features; the distribution samples include confused samples and known samples; An algorithm employing alternating boosting of classifier C and generator G guides classifier C to focus on known samples, correcting its bias of over-focusing on confused samples. This includes boosting the stochastic gradient to update the parameters of discriminator D. The parameters of generator G Parameters of classifier C , represented as: , , ; Minimize the overall loss to update the parameters of classifier C. Until convergence, it can be represented as: .

[0016] This invention employs an open-set recognition algorithm for wireless communication signals based on convolutional block attention and reciprocity points. By combining channel attention and spatial attention mechanisms, the feature extraction network focuses more on the energy accumulation features in the time-frequency plot of the wireless communication signal. Ultimately, this enables the classifier to better mine signal features, improving the accuracy of identifying abnormal and unknown signals in real electromagnetic environments. In particular, it enables the identification of abnormal and unknown signals in wireless communication fields with unknown performance, redundant information in the time-frequency plot, and under the influence of noise and interference in complex open electromagnetic environments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of an open-set identification method for wireless communication signals according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the time-frequency graph of a normal communication signal provided according to an embodiment of the present invention; Figure 3This is a time-frequency diagram of an abnormal communication signal provided according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the reciprocity point model structure provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the spatial distance between a given signal sample and a reciprocal point, provided according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a generative adversarial network model for generating obfuscated samples according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a feature extraction network architecture provided according to an embodiment of the present invention; Figure 8 This is a comparison chart of the recognition performance of open set recognition models under different signal-to-noise ratios according to embodiments of the present invention; Figure 9 This is a comparison chart of the recognition performance of open set recognition models under different interference-to-signal ratios provided in the embodiments of the present invention. Detailed Implementation

[0018] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] The wireless communication signal open set recognition method based on convolutional block attention and reciprocity points provided by this invention is as follows: Figure 1 As shown, it includes the following steps: Step S100: Receive signal data, which includes K types of known signal data and U types of unknown signal data; label the K types of known signal data to form a training dataset. The K-type known signal data and the U-type unknown signal data constitute an open test dataset; In practical applications, in addition to legitimate signals, malicious or illegal signals may also exist in the signal data received during communication. Some of these malicious or illegal signals may have been received before. Therefore, these malicious or illegal signals can be considered as one of the known categories. If this type of signal can be distinguished from other known categories, it can be labeled as an abnormal signal. Thus, known signal data includes both normal and abnormal wireless communication signals.

[0020] Define the training dataset in this step. Training dataset It contains training data with K known classes, represented as: ,in yes Tags; Define an open test dataset , represented as ,in The tags contain Where K represents the category of known classes in the real-world scenario. This represents the category of an unknown class in a real-world scenario; therefore, communication signal data in a complex open electromagnetic environment can be represented as: Its structure and Consistent, where k+1 to k+u represent unknown signals of class U.

[0021] Define the training dataset The feature space to which it belongs after processing by the feature extraction network is: Signal category The feature space of the dataset is used This indicates that the corresponding open space is In order to effectively formalize and manage the risks of open spaces, Divided into two subspaces: the positive open space corresponding to the known class signal is The remaining infinite unknown space is negative open space. In other words, .

[0022] Therefore, the open set identification that this invention needs to achieve, that is, minimizing the open space combination of labeled known class risks and unknown class risks within the allowed identification space R, is expressed as: (1) The training of the model involved in the subsequent steps of this invention, samples From known categories ,sample Samples from other known categories From Unknown categories other than those mentioned above. These three types of samples from different sources are defined as positive training data, negative training data, and potentially unknown data, respectively.

[0023] Step S110: Convert the signal data in the training dataset and the open test dataset into a time-frequency graph; In this step, smoothed pseudo-Wigner-Ville time-frequency analysis is used to transform the signal data in the training dataset and the open test dataset into a two-dimensional time-frequency plot, represented as: (2) in It is a time difference variable. Indicates time, Indicates frequency, Represents the signal function. and It is a real symmetric window function: Intersecting terms that oscillate parallel to the time axis are allowed to perform time smoothing. Frequency smoothing is performed on cross terms that oscillate parallel to the frequency axis.

[0024] Training dataset and open test dataset A schematic diagram of the time-frequency graph of a normal communication signal is shown below. Figure 2 As shown, the time-frequency diagram of the abnormal communication signal is as follows: Figure 3 As shown.

[0025] At this point, the training dataset and open test dataset All of these are signal data sets based on time-frequency diagrams.

[0026] In subsequent steps, the training dataset will be... The input is fed into a classification model based on convolutional block attention and reciprocity points for training; then the open test dataset is used. Inputting the data into a trained classification model will reveal the types of open data, including known types. Class and unknown U class.

[0027] Step S120: Construct a dataset based on the training dataset The reciprocal point model obtains the latent feature representation of unknown signal samples in the out-of-class space corresponding to each known class of signal data, and initializes them as random distributions.

[0028] Constructing a reciprocal point model includes the following steps: 1) First, consider the reciprocity points of wireless communication signal category K. Perform modeling; The structure of the reciprocal point model is as follows Figure 4 As shown, the reciprocity points of communication signal category K are... yes Out-of-class feature space Potential representation samples As can be seen from the layout of the points in the reciprocal point model, signal sample ratio The signal samples are closer to the reciprocity point That is, in the entire feature space, the distance between a reciprocal point and an unknown class is smaller than the distance between a reciprocal point and a known class.

[0029] 2) Furthermore, the distance between a given signal sample and a reciprocal point is defined using the reciprocal point model; In a given signal sample and reciprocal points Under the condition that the spatial distance between reciprocal points of signal categories is as follows: Figure 5 As shown, given signal samples The distance dist(A, B) between point A and reciprocal point B is determined by both Euclidean distance and cosine distance, and can be obtained through... 3D feature representation, expressed as: (3) in, This represents the feature extraction function, which is the function that maps signal samples to the feature space, where m is the dimension.

[0030] 3) Based on the distance between given signal samples and reciprocal points, perform recognition and classification learning to determine the known and unknown spaces: Based on the characteristics of reciprocal points, and using distance-based metrics, samples The probability of belonging to category K and With reciprocal points The differences between them are directly proportional, which shows that and The greater the distance between them, the better the sample Marked as category The higher the probability of identification, the greater the probability of identification, and the sum of the probabilities of all categories is 1. Therefore, we define the identification classification probability and normalize it using the softmax function, expressed as: (4) in, It is a hyperparameter that controls the distance-probability transformation.

[0031] The process of recognition and classification learning is achieved by minimizing the reciprocal point classification loss based on the negative log probability of the known class K, expressed as: (5) in These are the parameters of the feature extraction function. For classifying losses, To identify classification probabilities, As a feature, These are reciprocal points.

[0032] At this point, the known and unknown spaces can be separated by maximizing the distance between the reciprocal points of the signal categories and their corresponding training samples.

[0033] Step S130: Based on the training dataset A classification model is trained, which is based on an adversarial network model and a feature extraction network. First, the risk of open space for each known category is reduced by limiting open space to bounded areas.

[0034] In addition to classifying known classes, the classifier model separates the known and unknown spaces by maximizing the distance between the reciprocity points of signal classes and their corresponding training samples, as follows: (6) in This represents the classification function.

[0035] To further reduce the risks of unknown open spaces, [the following will be included] Samples and reciprocal points Distance constraints between The risk of limiting open spaces within a certain scope is represented as: (7) in It is a learnable boundary distance, initially set to 1. The distance between the sample and the reciprocal point is constrained in Maximum boundary value within the range, This indicates that Euclidean distance is used to obtain a larger range of non-K samples.

[0036] On the other hand, categories With other exchange points The boundary is represented as: (8) Through this "adversarial" mechanism between equations (6) and (8), each known signal class is pushed to the edge of the finite feature space to the maximum extent. Therefore, by generating confused samples as unknown signal data through the adversarial network model, the recognition accuracy of unknown signal classes that are close to the features of known signal classes can be improved.

[0037] An adversarial network model includes: a discriminator D, a generator G, and a classifier C; the discriminator D represents the probability that a sample comes from the true distribution or a false distribution, and the generator G generates a sample... From the prior distribution Mapped to the output, the classifier C represents the probability that a sample belongs to each known class. Given a prior distribution... Generate samples and known signal samples The discriminator D is optimized to distinguish between real and generated fake samples.

[0038] Specifically, the definition of an adversarial network model includes the following steps: 1) Generate obfuscated samples: The reciprocity point dataset is input into the generator G to generate confused samples. The generative adversarial network model of the confused samples is as follows: Figure 6 As shown; 2) Optimize the discriminator: The confused samples are input into the discriminator D for discrimination, and these samples are identified as known class samples; the classifier C extracts the features of the signal samples and compares them with the reciprocity point dataset to classify the samples in the training dataset into known classes, and calculates the loss of the discriminator D: (9) in This represents the probability of identifying a known class of signal samples as a real sample. This represents the probability of identifying a generated confused sample as a fake sample.

[0039] The generator G aims to generate samples that are closer to the known class, thereby deceiving the discriminator D. The loss of the generator G is calculated as follows: (10) To obfuscate generators, an adversarial mechanism is introduced between known classes and reciprocal points. This encourages generators to operate at each center of the open space. Creating samples in the vicinity is equivalent to encouraging the generator to produce images that approximate the global open space. Formally, the generator is optimized through a classifier, represented as: (11) in .

[0040] The loss reaches its maximum when the features of the confused samples are close to all reciprocal points. The final loss of the generator G is calculated as follows: (12) (13) in, To control the hyperparameters of the information entropy loss weights, This represents the information entropy loss function.

[0041] The loss of classifier C is optimized using the generated confused samples: (14) in This is the overall loss function.

[0042] 3) Joint training: It is important to note that the known samples and generated samples are processed independently in Equation (14). In this case, since the distributions of the generated samples and the known samples are different, the classifier C may be confused, thus affecting the accuracy of the statistical results. To decompose the mixed distribution samples into two basic distributions, known samples and confused samples, an Auxiliary Batch Normalization (ABN) method is proposed to ensure that normalized statistics are obtained for the confused samples. Specifically, batch normalization standardizes the input features by calculating the mean and variance of each mini-batch. At this time, the input features should come from a single or similar distribution. Auxiliary batch normalization effectively decouples the mixed distribution samples by establishing independent batch normalization statistics for features of different domains.

[0043] Finally, through joint training, an algorithm that alternately improves classifier C and generator G, classifier C is guided to focus on known samples, correcting its bias of over-focusing on confused samples. (15) (16) (17) The parameters of the discriminator D are updated by boosting the stochastic gradient using equations (15), (16), and (17). The parameters of generator G Parameters of classifier C After updating these three parameters, focus training is added, that is, the classifier C is retrained using known samples, and the parameters of the classifier C are updated by minimizing the overall loss in equation (18). Until convergence: (18) This joint training aims to guide classifier C to focus on known samples, correcting its bias of over-focusing on confused samples.

[0044] During the adversarial learning phase at reciprocal points It was set to 1.0. and The values ​​were set to 0.1, and they were all determined through cross-validation. The reciprocal point was initialized using a random normal distribution, with a point randomly selected as the reciprocal point, and each boundary distance... Initialize all values ​​to 1. The model training uses the SGD optimizer with a learning rate of 0.1. The learning rate of the classifier starts at 0.1 and decreases by a factor of 0.1 every 30 epochs during training. The epoch is 100. After training is complete, the classifier model is saved.

[0045] In addition, a feature extraction network is built, and a convolutional block attention module is introduced into the feature extraction network to make it pay more attention to the feature information of areas where energy is focused (i.e., where the signal exists) in the time-frequency graph. The architecture of the convolutional block attention module is as follows: Figure 7 As shown, the network is mainly composed of nine cascaded convolutional modules with a kernel size of 3×3. Each convolutional module has a batch normalization (BN) layer after its convolutional layers to normalize all data, thereby improving training efficiency, before finally applying the ReLU activation function. To prevent overfitting, a dropout layer is added to the network.

[0046] Furthermore, a Convolutional Block Attention (CBAM) module is added to the feature extraction network. Specifically, the features extracted from the nine convolutional layers are normalized, then processed by the channel attention module, and the result is further processed by the spatial attention module. Given a set of feature maps... As input, CBAM sequentially infers a one-dimensional channel attention map. and 2D spatial attention map The overall attention process can be summarized as follows: (19) in, For element-wise multiplication, This is the feature map after processing by the channel attention module. This is the feature map after processing by the spatial attention module.

[0047] The implementation of channel attention includes: firstly, aggregating the spatial information of the feature map through max pooling and average pooling to obtain... and The results from the two pooling processes are then forwarded to a shared network for learning. This shared network consists of a multilayer perceptron. Finally, the results from the shared network are summed element-wise and merged to output a feature vector. This vector is then processed by the Sigmoid function to obtain the final channel attention value.

[0048] The implementation of spatial attention involves: firstly, aggregating the channel information of the feature map through max pooling and average pooling operations to generate two two-dimensional maps: and The two 2D graphs are then concatenated and convolved through a standard convolutional layer to generate a 2D spatial attention map.

[0049] Finally, the feature vector is obtained by passing through the ReLU activation function, average pooling layer, and fully connected layer.

[0050] At this point, the classification model consisting of the adversarial network model and the feature extraction network is defined.

[0051] In this step, the training dataset is used. Training a classification model, where the training dataset Model training only requires a dataset containing K known classes of wireless communication signals.

[0052] Step S140: Open the test dataset The input is fed into a trained classification model, which outputs K+1 class labels for the signal class corresponding to the maximum reciprocity distance, where K is the number of known signal classes. "All of them are identified as unknown category labels, that is, subclasses of the K+1th unknown class."

[0053] In the implementation of open-set recognition of wireless communication signals, the time-frequency graph has the unique characteristic of redundant background information: that is, the original open-set recognition model based on reciprocity points exhibits poor performance when migrated to the wireless communication domain. The open-set recognition method for wireless communication signals provided in this invention adds a convolutional block attention module to the feature extraction network, making it pay more attention to the feature information of signal energy accumulation in the time-frequency graph, thereby improving the classification accuracy of known signal classes and better detecting unknown signal classes, such as... Figure 8 and Figure 9 As shown, even under conditions of low signal-to-noise ratio (SNR) and low interference-to-signal ratio (JSR), a comparison of the recognition performance of the open set recognition model based on convolutional block attention and reciprocity points proposed in this invention with the original open set recognition model based on reciprocity points clearly demonstrates the beneficial effect of the method described in this invention on open set recognition of wireless communication signals.

[0054] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for open set recognition of wireless communication signals based on convolutional block attention and reciprocity points, characterized in that, Includes the following steps: Receive wireless communication signal data, which includes K types of known signal data and U types of unknown signal data; label the K types of known signal data to form a training dataset. The K-type known signal data and the U-type unknown signal data constitute an open test dataset. The training dataset The feature space to which it belongs is Signal category The feature space of the dataset is used This indicates that the corresponding open space is ; The training dataset and open test dataset The signal data in the diagram is converted into a time-frequency graph; Construct based on the training dataset The reciprocity point model is used to obtain the latent feature representation of unknown signal samples in the out-of-class space corresponding to each known class of signal data; Based on the training dataset A classification model is trained, which is based on an adversarial network model and a feature extraction network. The open test dataset The input is fed into the classification model, which outputs the K+U class labels of the signal class corresponding to the maximum reciprocity distance, where K is the number of known signal classes. "To identify categories that are classified as unknown." 2. The open-set identification method for wireless communication signals according to claim 1, characterized in that, The known signal data includes legitimate signals and malicious / illegal signals; legitimate signals are labeled as normal wireless communication signals; and malicious / illegal signals are labeled as abnormal wireless communication signals.

3. The open-set identification method for wireless communication signals according to claim 2, characterized in that, The training dataset It contains training data with K known classes, represented as: ,in yes Tags; The open test dataset Represented as: , where k+1 to k+u represent unknown signals of class U.

4. The open-set identification method for wireless communication signals according to claim 1, characterized in that, The conversion to a time-frequency graph is achieved using smoothed pseudo-Wigner-Ville time-frequency analysis, and is represented as follows: ， in, It is a time difference variable. Represents a time variable. Represents frequency variables. Represents the signal function. Allows cross terms that oscillate parallel to the time axis to execute a time-smoothed real symmetric window function. Cross terms that oscillate parallel to the frequency axis are allowed to perform frequency-smoothed real symmetric window functions.

5. The open-set identification method for wireless communication signals according to claim 1, characterized in that, The construction is based on the training dataset. The reciprocity point model includes the following steps: Reciprocity points for wireless communication signal class K Modeling is performed where the reciprocity points of communication signal category K are... yes Out-of-class feature space Potential representation samples ; The distance between a given signal sample and a reciprocal point is defined using a trading point model. Based on the distance between a given signal sample and a reciprocal point, identification and classification learning is performed to determine the known and unknown spaces.

6. The open-set identification method for wireless communication signals according to claim 5, characterized in that, The distance between the given signal sample and the reciprocal point is determined by both the Euclidean distance and the cosine distance, expressed as: ， in, This represents the distance between a given signal sample and a reciprocal point. This represents the feature extraction function, which is the function that maps signal samples to the feature space, where m is the dimension.

7. The open-set identification method for wireless communication signals according to claim 5, characterized in that, The recognition and classification learning includes: Define the classification probability and normalize it using the softmax function, expressed as: ,in, It is a hyperparameter that controls the distance-probability transformation; Recognition and classification learning is achieved by minimizing the reciprocal point classification loss based on the known negative log probabilities of class K, expressed as: ,in These are the parameters of the feature extraction function.

8. The open-set identification method for wireless communication signals according to claim 1, characterized in that, Based on the training dataset Training a classification model includes the following steps: will come from Samples and reciprocal points Distance constraints between scope; A classification model is defined, which includes an adversarial network model and a feature extraction network; the adversarial network model generates confused samples as unknown signal data; the feature extraction network introduces a convolutional block attention module to make it pay more attention to the feature information of the energy focus area in the time-frequency map. Using training dataset Train the classification model.

9. The open-set identification method for wireless communication signals according to claim 8, characterized in that, The adversarial network model includes a discriminator D, a generator G, and a classifier C; The discriminator D represents the probability that a sample comes from the true distribution or the false distribution, and the generator G generates a sample. From the prior distribution Mapped to the output, the classifier C represents the probability that a sample belongs to each known class. Given a prior distribution... Generate samples and known signal samples The discriminator D is optimized to distinguish between real and generated fake samples.

10. The open-set identification method for wireless communication signals according to claim 9, characterized in that, The adversarial network model is generated through joint training, including the following steps: Mixed distribution samples are decoupled by auxiliary batch normalization; the auxiliary batch normalization refers to establishing independent batch normalization statistics for different domain features; the distribution samples include confused samples and known samples; An algorithm employing alternating boosting of classifier C and generator G guides classifier C to focus on known samples, correcting its bias of over-focusing on confused samples. This includes boosting the stochastic gradient to update the parameters of discriminator D. The parameters of generator G Parameters of classifier C , represented as: ， ， ； Minimize the overall loss to update the parameters of classifier C. Until convergence, it can be represented as: 。

Images