Unmanned aerial vehicle radio frequency signal identification method, equipment and medium

Abstract

The invention relates to an unmanned aerial vehicle radio frequency signal identification method and device, and a medium, and the method comprises the steps: collecting to-be-identified unmanned aerial vehicle radio frequency IQ data, and carrying out the short-time Fourier transform of the to-be-identified unmanned aerial vehicle radio frequency IQ data, and obtaining an unmanned aerial vehicle time-frequency graph; inputting the to-be-identified unmanned aerial vehicle radio frequency IQ data and the unmanned aerial vehicle time-frequency graph into an open set unmanned aerial vehicle radio frequency signal identification model to obtain the type of the unmanned aerial vehicle; wherein the open set unmanned aerial vehicle radio frequency signal identification model comprises an image encoder module, a semantic alignment module and a class decorrelation module, and the image encoder module comprises a flight control signal block extractor part and a flight control signal mode extractor part. According to the invention, the detection capability of the unmanned aerial vehicle of the unknown category under the open set condition is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle detection, and in particular to an unmanned aerial vehicle radio frequency signal identification method, device and medium. BACKGROUND

[0002] In recent years, the rapid development of unmanned aerial vehicles (UAVs) has brought significant benefits to many fields such as agricultural irrigation, power inspection, geographic mapping, and communication relay. However, as the application scenarios of unmanned aerial vehicles continue to expand, their misuse has raised public concerns about safety and privacy. Therefore, efficient and reliable detection and identification of illegal unmanned aerial vehicles that may threaten the safety of critical infrastructure have become a pressing problem.

[0003] The detection and identification technology for unmanned aerial vehicles mainly includes radar, vision, acoustic, and radio frequency (RF) monitoring. Compared with radar, vision, and acoustic methods, passive methods based on radio frequency have become an important means of monitoring unmanned aerial vehicles in urban environments due to their advantages such as not being limited by line of sight, strong anti-Non Line of Sight (NLOS) capability, and being friendly to urban environments. Existing unmanned aerial vehicle radio frequency signal identification methods usually identify and locate by monitoring the video transmission signal (VTS) of the unmanned aerial vehicle, but this method requires the unmanned aerial vehicle to keep the video transmission signal continuously open. However, the video transmission signal can be manually turned off, and in addition, the video transmission signal is located in the 5.8G frequency band, and the signal attenuates severely with increasing distance, making it easy for illegal unmanned aerial vehicles to evade detection by the radio frequency monitoring system.

[0004] To address the above challenges, more and more research has begun to focus on the more basic flight control signal (FCS) of unmanned aerial vehicles. Flight control signals bear the routine flight control instructions and log transmission functions between unmanned aerial vehicles and ground control terminals, and are usually modulated using strong anti-interference technologies such as frequency hopping spread signal (FHSS), making them non-closable and thus more stable and non-avoidable than video transmission signals. However, most current FCS-based detection methods still assume that the training and test signals belong to a closed set, leading to unknown types of signals being easily misjudged as known categories, and making it impossible to cope with the explosive growth of unknown unmanned aerial vehicle signals in an open set environment. SUMMARY

[0005] The technical problem solved by the present application is to provide a UAV radio frequency signal identification method, device and medium, which improves the detection capability of unknown category UAVs under open set conditions.

[0006] The technical solution adopted by the present application to solve its technical problem is to provide a UAV radio frequency signal identification method, comprising the following steps:

[0007] Collecting the UAV radio frequency IQ data to be identified, and performing short-time Fourier transform on the UAV radio frequency IQ data to be identified to obtain a UAV time-frequency graph;

[0008] Inputting the UAV radio frequency IQ data to be identified and the UAV time-frequency graph into an open set UAV radio frequency signal identification model to obtain the category of the UAV;

[0009] The open set UAV radio frequency signal identification model comprises:

[0010] An image encoder module is configured to obtain the category prediction probability and image semantic features of the UAV according to the UAV radio frequency IQ data to be identified and the UAV time-frequency graph;

[0011] A semantic alignment module is configured to calculate a semantic alignment loss according to the image semantic features and text semantic features to promote the alignment of images and texts in the semantic space;

[0012] A class decorrelation module is configured to calculate a class decorrelation loss according to the category representative features to improve the distinguishability between the category features.

[0013] The image encoder module comprises:

[0014] A flight control signal block extractor part is configured to extract detailed edge and multi-scale texture features from the UAV time-frequency graph;

[0015] A flight control signal mode extractor part is configured to extract overall frequency hopping features from the UAV radio frequency IQ data to be identified;

[0016] A global pooling part is configured to complete the classification of the UAV according to the detailed edge and multi-scale texture features.

[0017] The flight control signal block extractor part comprises a first progressive residual convolution unit, a Sobel edge feature enhancement branch unit, an SE channel attention unit, a second progressive residual convolution unit, a third progressive residual convolution unit and a fourth progressive residual convolution unit, the input of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit is the UAV time-frequency graph, and the output of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit is spliced as the input of the SE channel attention unit; the SE channel attention unit, the second progressive residual convolution unit, the third progressive residual convolution unit and the fourth progressive residual convolution unit are connected in sequence.

[0018] The first progressive residual convolution unit, the second progressive residual convolution unit, the third progressive residual convolution unit and the fourth progressive residual convolution unit have the same structure, and each comprises a first convolution layer, a second convolution layer, a third convolution layer and a fourth convolution layer; the first convolution layer and the second convolution layer are connected in cascade, the third convolution layer and the fourth convolution layer are connected in cascade, the output of the first convolution layer and the output of the second convolution layer are spliced as the input of the third convolution layer; and the output of the second convolution layer and the output of the fourth convolution layer are added as the output of the whole convolution unit.

[0019] The flight control signal mode extractor part comprises a two-dimensional position coding unit and a self-attention unit, the input of the two-dimensional position coding unit is the UAV radio frequency IQ data to be identified, and the output is added to the UAV radio frequency IQ data to be identified as the input of the self-attention unit.

[0020] The calculation method of the semantic alignment loss is as follows: Wherein, L SC is the semantic alignment loss, CE() is the cross-entropy loss function, I represents the image semantic feature, T represents the text semantic feature, tau is the temperature parameter, and labels represent the unit matrix label with the diagonal line being the correct matching pair.

[0021] The calculation method of the class decorrelation loss is as follows: L De = CE(M, diag(s)), wherein L De is the class decorrelation loss, CE() is the cross-entropy loss function, M is the cosine similarity matrix between the class representative features, and diag(s) is the diagonal line label of the unit matrix.

[0022] In the verification phase of the open set unmanned aerial vehicle radio frequency signal identification model, whether a sample is an unknown category is determined by an adaptive entropy threshold method, and the calculation method of the adaptive entropy threshold is: τ'= μ1+ α' σ1, wherein τ' is the adaptive entropy threshold, μ1 and σ1 are respectively the mean and standard deviation of the known class Gaussian distribution, α' is the standard deviation multiple, and is expressed as: a, b are respectively the lower limit and upper limit of the Sigmoid function, k is the steepness of the curve, and α is a discrimination coefficient, and is expressed as: μ2 is the mean of the unknown class Gaussian distribution, and x0 is the midpoint position of the discrimination coefficient. a, b are respectively the lower limit and upper limit of the Sigmoid function, k is the steepness of the curve, and α is a discrimination coefficient, and is expressed as: μ2 is the mean of the unknown class Gaussian distribution, and x0 is the midpoint position of the discrimination coefficient. μ2 is the mean of the unknown class Gaussian distribution, and x0 is the midpoint position of the discrimination coefficient.

[0023] The technical scheme adopted by the present application to solve its technical problems is to provide an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the unmanned aerial vehicle radio frequency signal identification method.

[0024] The technical scheme adopted by the present application to solve its technical problems is to provide a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the unmanned aerial vehicle radio frequency signal identification method.

[0025] Advantages

[0026] Compared with the prior art, the present application has the following advantages and positive effects: the present application designs a lightweight open set unmanned aerial vehicle radio frequency signal identification model, which can effectively extract the features of the flight control signal through the flight control signal block extractor and the flight control signal mode extractor, and generate an explicit semantic anchor point for each known category through the semantic alignment module, thereby reducing the phenomenon that the model is wrongly judged as a known category with similar semantics when encountering an unknown category, and encouraging the separation between known clusters through the class decorrelation module, effectively enhancing the distance between clusters, thereby promoting better separation of known categories, and improving the detection capability of unknown unmanned aerial vehicles under open set conditions. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flowchart of the unmanned aerial vehicle radio frequency signal identification method of the first embodiment of the present application;

[0028] Figure 2 is a framework diagram of the open set unmanned aerial vehicle radio frequency signal identification model in the first embodiment of the present application;

[0029] Figure 3 is a structural diagram of the progressive residual convolution unit in the first embodiment of the present application. DETAILED DESCRIPTION

[0030] The application will be further described in connection with the specific embodiments. It should be understood that these embodiments are only used to illustrate but not to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content of the application, and these equivalent forms also fall within the scope defined by the appended claims.

[0031] The first embodiment of the application relates to a method for identifying radio frequency signals of unmanned aerial vehicles, as shown in the figure, comprising the following steps: Figure 1

[0032] Step 1, collecting the radio frequency IQ data of the unmanned aerial vehicle to be identified, and performing short-time Fourier transform on the radio frequency IQ data of the unmanned aerial vehicle to be identified to obtain a time-frequency diagram of the unmanned aerial vehicle.

[0033] Step 2, inputting the radio frequency IQ data of the unmanned aerial vehicle to be identified and the time-frequency diagram of the unmanned aerial vehicle into an open set radio frequency signal identification model of the unmanned aerial vehicle to obtain the category of the unmanned aerial vehicle.

[0034] As shown in the figure, the open set radio frequency signal identification model of the unmanned aerial vehicle in the embodiment comprises an image encoder module, a semantic alignment module and a class decorrelation module, wherein the image encoder module comprises a flight control signal block extractor part and a flight control signal pattern extractor part. Figure 2

[0035] The flight control signal block extractor part is used to extract detailed edges and multi-scale texture features from the time-frequency diagram of the unmanned aerial vehicle, and has strong identification ability especially for flight control signals with obvious frequency hopping. The flight control signal block extractor part in the embodiment can efficiently and accurately capture the key spectral features of the flight control signal of the unmanned aerial vehicle, and is especially suitable for spectral identification tasks in complex electromagnetic environments. In order to achieve this goal, the flight control signal block extractor part comprises a first progressive residual convolution unit, a Sobel edge feature enhancement branch unit, an SE channel attention unit, a second progressive residual convolution unit, a third progressive residual convolution unit and a fourth progressive residual convolution unit, the inputs of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit are the time-frequency diagram of the unmanned aerial vehicle, and the outputs of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit are spliced as the input of the SE channel attention unit; the SE channel attention unit, the second progressive residual convolution unit, the third progressive residual convolution unit and the fourth progressive residual convolution unit are connected in sequence. The first progressive residual convolution unit, the second progressive residual convolution unit, the third progressive residual convolution unit and the fourth progressive residual convolution unit have the same structure.

[0036] The main branch feature extraction of the flight control signal block extractor part adopts a progressive residual convolution (PRC) unit, as shown in the figure.​​Figure 3 As shown, the PRC unit includes a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth convolutional layer; the first convolutional layer and the second convolutional layer are cascaded, the third convolutional layer and the fourth convolutional layer are cascaded, the output of the first convolutional layer and the output of the second convolutional layer are spliced to serve as the input of the third convolutional layer; the output of the second convolutional layer and the output of the fourth convolutional layer are added to serve as the output of the entire convolutional unit. As can be seen, the PRC unit gradually extracts and enhances the fine features of the time-frequency graph of the unmanned aerial vehicle through convolution, channel splicing, and residual connection operations, effectively alleviating the gradient vanishing problem of the network.

[0037] The specific implementation process of the PRC module is as follows: first, the input features are respectively extracted by two consecutive convolutional layers, and the convolution kernel size is 3x3, and appropriate padding is used to ensure that the spatial size of the feature map remains unchanged. Subsequently, the feature maps obtained by two times of convolution are spliced along the channel dimension to generate a feature map with more rich and detailed feature expression. Then, the spliced feature map is subjected to batch normalization processing, and then subjected to two consecutive convolutional layers (each convolution is accompanied by ReLU activation), and added to the feature map before convolution through residual connection (i.e. shortcut connection) to further enhance the feature expression ability and stability of the network. The main branch feature undergoes progressive convolution and feature enhancement, and outputs a high-dimensional and detailed feature expression.

[0038] The flight control signal mode extractor part in the embodiment is used to extract the overall frequency hopping feature from the to-be-identified unmanned aerial vehicle radio frequency IQ data. The flight control signal mode extractor part includes a two-dimensional position encoding unit and a self-attention unit, the input of the two-dimensional position encoding unit is the to-be-identified unmanned aerial vehicle radio frequency IQ data, and the output is added to the to-be-identified unmanned aerial vehicle radio frequency IQ data to serve as the input of the self-attention unit. The flight control signal mode extractor part is a position-aware global feature extractor, which forms an effective complement with the flight control signal block extractor part and is specially used to capture the overall frequency hopping in the unmanned aerial vehicle communication signal.

[0039] The semantic alignment module in the embodiment is used to calculate a semantic alignment loss according to the image semantic features and the text semantic features, so as to promote the alignment of the image and the text in the semantic space. The semantic alignment module introduces a text end encoder of a CLIP pre-training model as a semantic anchor generator, provides an explicit semantic representation for each known category, and then realizes the accurate alignment of the image semantic features and the text semantic features in the semantic space through a bidirectional semantic cross-entropy loss function, thereby effectively solving the problem that a traditional clustering manner ignores an explicit semantic boundary. Through such an explicit semantic anchoring, the embodiment significantly improves the ability of the model to distinguish between unknown categories and known categories that are similar in semantics in an open set scenario, and has a unique semantic distinguishing advantage. The semantic alignment loss adopts a bidirectional cross-entropy loss, and the calculation manner is as follows:

[0040]

[0041] wherein, L SC is the semantic alignment loss, CE() is a cross-entropy loss function, I represents image semantic features, T represents text semantic features, τ is a temperature parameter for adjusting the sensitivity of feature similarity, and labels represent a unit matrix label with a correct matching pair on the diagonal. The semantic alignment loss promotes the mutual aggregation of semantic-related features in the feature space and improves the ability of unknown category detection.

[0042] The category decorrelation module in the embodiment is used to calculate a category decorrelation loss according to the category representative features, so as to improve the distinguishability between category features. The category decorrelation module proposes a dynamic cosine similarity matrix optimization mechanism based on the category representative features by using the orthogonality constraint between features, significantly improves the feature separation degree between known categories, and thereby effectively suppresses the misjudgment rate between known categories and unknown categories in an open set recognition task. In the embodiment, the batch random selection category representative feature method is adopted, the category representative features are dynamically constructed in each batch training, thereby enhancing the adaptability of the model to the changes in the feature space, the cosine similarity matrix calculation method is used, and further the feature decorrelation constraint between categories is realized through Softmax regularization. The calculation manner of the category decorrelation loss is as follows:

[0043] L De = CE(M, diag(s));

[0044] wherein, L De is the category decorrelation loss, M is a cosine similarity matrix between the category representative features, and diag(s) is a unit matrix diagonal line label. The category decorrelation module guides the feature matrix to approach a unit matrix (the diagonal line elements are close to 1, and the non-diagonal line elements are close to 0), effectively reduces the feature correlation between categories, and significantly improves the generalization performance in an open set environment.

[0045] The open set UAV radio frequency signal recognition model in the embodiment adopts an end-to-end training strategy, and simultaneously optimizes classification accuracy, semantic feature alignment, and inter-class discrimination. The specific training process is as follows:

[0046] (1) Randomly batch the training data, and each batch of data contains a UAV spectrum image and corresponding text embedding features and class labels;

[0047] (2) Input the spectrum image into the image encoder, and output the class prediction probability and image semantic features;

[0048] (3) Input the text embedding features into the text re-encoder, and output the normalized text semantic features;

[0049] (4) Calculate and combine three loss functions:

[0050] The classification cross-entropy loss L CE ensures that the model classifies correctly;

[0051] The semantic alignment loss L SC promotes the alignment of the image and text feature spaces;

[0052] The class decorrelation loss L De improves the discrimination between class features;

[0053] Combine the above losses with weight coefficients (η1, η2, η3) to form the total loss L, that is, L = η1L CE + η2L SC + η3L De ;

[0054] (6) Calculate the gradient using the back propagation algorithm to update the network parameters;

[0055] (7) In the verification stage, determine whether the sample is of an unknown class by using the adaptive entropy threshold method. Calculate the entropy of the classification output of each sample. When the entropy value exceeds the set threshold, the sample is determined to be of an unknown class, otherwise it is determined to be of a known class;

[0056] In order to accurately distinguish between known and unknown classes in an open set environment, an adaptive entropy threshold method based on Gaussian Mixture Model (GMM) and Sigmoid dynamic mapping is used. This method aims to adaptively determine a dynamic classification entropy threshold according to the uncertainty of the model's classification of the sample (entropy value) to optimize the division accuracy of known and unknown classes. The specific implementation process is described as follows:

[0057] First, for the classification probability output by the model, define the information entropy (Entropy) as a measure of the model's prediction uncertainty for a certain sample:

[0058]

[0059] where H(x) is the classification entropy of sample x, p c (x) is the probability of sample x being predicted as class c by the model; C is the total number of classes. The higher the entropy value, the more uncertain the model is about the predicted class of the current sample.

[0060] After obtaining the entropy values of a large number of samples, the entropy distribution is fitted by a Gaussian Mixture Model (GMM). The GMM assumes that the overall entropy distribution can be approximated by the weighted superposition of two Gaussian distributions, one corresponding to the low-entropy region (usually the known class) and the other corresponding to the high-entropy region (usually the unknown class). The fitted entropy distribution can be represented as:

[0061]

[0062] where μ1,σ1 are the mean and standard deviation of the Gaussian distribution of the low-entropy region (known class); μ2,σ2 are the mean and standard deviation of the Gaussian distribution of the high-entropy region (unknown class); π1,π2 are the weight coefficients of the two Gaussian distributions, satisfying π1+π2=1.

[0063] The specific parameters of the two Gaussian distributions are determined by optimizing the above parameters through the Expectation-Maximization algorithm (EM).

[0064] According to the two fitted Gaussian distributions, the ratio of the distance between the means of the two distributions to the standard deviation of the low-entropy distribution is defined as the discrimination coefficient α:

[0065]

[0066] The larger the discrimination coefficient α, the more obvious the distinction between the two Gaussian distributions, i.e., the higher the entropy distinction between the known class and the unknown class. Further, a dynamic Sigmoid function is defined to determine the specific standard deviation multiple α':

[0067]

[0068] where parameters a, b are the lower and upper limits of the Sigmoid function, respectively; parameter k represents the steepness of the curve; parameter x0 represents the midpoint position of the discrimination coefficient.

[0069] In this way, α' can dynamically adjust according to the change of the discrimination coefficient α, making the threshold more adaptive.

[0070] Based on the above dynamic mapping method, the final adaptive entropy threshold τ' is calculated as:

[0071] τ' = μ1 + α' σ1

[0072] When the entropy value of a certain test sample exceeds the threshold value, it is identified as an unknown class, otherwise it is determined as a known class.

[0073] (8) During training, record the loss changes of the training set and the validation set, and monitor the validation accuracy. When the performance of the validation set no longer improves, the early stopping strategy is adopted to prevent overfitting.

[0074] (9) Finally, save the best model parameters after training.

[0075] It can be found that the open set unmanned aerial vehicle radio frequency signal recognition model is designed, the feature of the flight control signal can be effectively extracted through the flight control signal block extractor and the flight control signal mode extractor, a clear semantic anchor point is generated for each known class through the semantic alignment module, the phenomenon that the model is wrongly judged as a known class with similar semantics when encountering an unknown class is weakened, the separation between known clusters is encouraged through the class decorrelation module, the distance between clusters is effectively enhanced, and therefore the better separation of the known classes is promoted, and the detection capability of the unknown unmanned aerial vehicle under the open set condition is improved.

[0076] The second embodiment of the present application relates to an electronic device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the unmanned aerial vehicle radio frequency signal recognition method of the first embodiment when executing the computer program.

[0077] The third embodiment of the present application relates to a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the unmanned aerial vehicle radio frequency signal recognition method of the first embodiment.

[0078] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in 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, etc.) containing computer usable program code.

[0079] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0080] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0081] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks. Figure 1 one or more flow or blocks.

[0082] The above description is only specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered by the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A method for identifying radio frequency signals of a UAV, characterized in that, The method comprises the following steps: Collecting the to-be-identified unmanned aerial vehicle radio frequency IQ data, and performing short-time Fourier transform on the to-be-identified unmanned aerial vehicle radio frequency IQ data to obtain an unmanned aerial vehicle time-frequency diagram; Inputting the to-be-identified unmanned aerial vehicle radio frequency IQ data and the unmanned aerial vehicle time-frequency diagram into an open set unmanned aerial vehicle radio frequency signal identification model to obtain the category of the unmanned aerial vehicle; The open set unmanned aerial vehicle radio frequency signal identification model comprises: An image encoder module configured to obtain category prediction probability and image semantic features of the unmanned aerial vehicle based on the to-be-identified unmanned aerial vehicle radio frequency IQ data and the unmanned aerial vehicle time-frequency diagram; A semantic alignment module configured to calculate a semantic alignment loss based on the image semantic features and text semantic features to promote the alignment of images and texts in a semantic space; A class decorrelation module configured to calculate a class decorrelation loss based on category representative features to improve the distinguishability between category features. 2.The UAV radio frequency signal identification method of claim 1, wherein, The image encoder module comprises: A flight control signal block extractor part configured to extract detailed edge and multi-scale texture features from the unmanned aerial vehicle time-frequency diagram; a flight control signal pattern extractor part configured to extract overall frequency hopping features from the to-be-identified unmanned aerial vehicle radio frequency IQ data; A global pooling part configured to complete the classification of the unmanned aerial vehicle based on the detailed edge and multi-scale texture features.

3. The method of claim 2, wherein, The flight control signal block extractor part comprises a first progressive residual convolution unit, a Sobel edge feature enhancement branch unit, an SE channel attention unit, a second progressive residual convolution unit, a third progressive residual convolution unit, and a fourth progressive residual convolution unit. The input of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit is the unmanned aerial vehicle time-frequency diagram. The output of the first progressive residual convolution unit and the Sobel edge feature enhancement branch unit is spliced as the input of the SE channel attention unit. The SE channel attention unit, the second progressive residual convolution unit, the third progressive residual convolution unit, and the fourth progressive residual convolution unit are connected in sequence.

4. The method of claim 3, wherein, The first progressive residual convolution unit, the second progressive residual convolution unit, the third progressive residual convolution unit, and the fourth progressive residual convolution unit have the same structure and comprise a first convolution layer, a second convolution layer, a third convolution layer, and a fourth convolution layer. The first convolution layer and the second convolution layer are cascaded, the third convolution layer and the fourth convolution layer are cascaded, and the output of the first convolution layer and the output of the second convolution layer are spliced as the input of the third convolution layer. The output of the second convolution layer and the output of the fourth convolution layer are added as the output of the entire convolution unit.

5. The method of claim 2, wherein, The flight control signal pattern extractor part comprises a two-dimensional position encoding unit and a self-attention unit. The input of the two-dimensional position encoding unit is the to-be-identified unmanned aerial vehicle radio frequency IQ data, and the output is added to the to-be-identified unmanned aerial vehicle radio frequency IQ data as the input of the self-attention unit.

6. The method of claim 1, wherein, The semantic alignment loss is calculated in the following manner: wherein L SC is the semantic alignment loss, CE() is a cross-entropy loss function, I represents image semantic features, T represents text semantic features, τ is a temperature parameter, and labels represents a unit matrix label with the correct matching pair on the diagonal.

7. The method of claim 1, wherein, The calculation manner of the quasi-decorrelation loss is: L De = CE(M, diag(s)), wherein, L De is the quasi-decorrelation loss, CE() is a cross-entropy loss function, M is a cosine similarity matrix between class representative features, and diag(s) is a unit matrix diagonal line label.

8. The drone radio frequency signal identification method of claim 1, wherein, In the verification phase of the open set unmanned aerial vehicle radio frequency signal identification model, whether the sample is of an unknown category is determined by an adaptive entropy threshold method, and the calculation method of the adaptive entropy threshold is τ'=μ 1 +α'σ 1, wherein τ' is the adaptive entropy threshold, μ 1 and σ 1 are respectively the mean and the standard deviation of the known class Gaussian distribution, and α' is a standard deviation multiple, and is represented as: a and b are respectively the lower limit and the upper limit of the Sigmoid function, k is the steepness of the curve, and α is a discrimination coefficient, and is represented as: μ2is the mean of the unknown Gaussian distribution, and x0is the midpoint position of the discriminant coefficients.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the unmanned aerial vehicle radio frequency signal identification method according to any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the method for identifying radio frequency signals of a UAV according to any one of claims 1-8.

Images