A method for radio frequency identification of unmanned aerial vehicles based on a deep attention detection model

Abstract

This invention discloses a drone radio frequency identification method based on a deep attention detection model, comprising: acquiring multiple drone radio frequency signals to be identified; inputting the multiple drone radio frequency signals to be identified into a pre-trained deep attention detection model, and predicting the classification of the output radio frequency signals; the radio frequency signal classification includes whether a drone exists, the drone model, and the current drone operating mode.

Description

Technical Field

[0001] This invention relates to a method for radio frequency identification (RFID) of unmanned aerial vehicles (UAVs) based on a deep attention detection model, belonging to the field of UAV RFID technology. Background Technology

[0002] With the application of advanced sensors such as GPS, LiDAR, radar, and optical cameras to drones, their use in surveying, surveillance, emergency rescue, forest fire fighting, entertainment, and military fields is becoming increasingly widespread. Since 2018, policies have been successively introduced to regulate drone flight management, but "black flight" incidents continue to occur, posing a significant threat to public safety. Drone detection and identification technology has emerged as a response to this need.

[0003] Drone detection and identification technologies include optical, acoustic, radio, and radar detection. Radar detection is an active detection method with a long detection range, but it is limited by the radar elevation angle and is prone to misidentification when detecting low-altitude, slow-moving, or small drones or flying creatures such as birds. Optical detection uses high-definition optical sensors for image tracking, providing good visualization, but the sensor's detection capability depends on weather conditions. In adverse weather conditions such as darkness, rain, and fog, the detection range is short, efficiency is poor, and there is a risk of malfunction. Acoustic detection captures drone sounds through acoustic sensors and has a certain recognition capability. It works well in quiet environments but is usually ineffective in crowded urban areas or noisy environments. Radio detection is mainly divided into two types: passive micro-Doppler detection and direct reception and identification of drone radio frequency signals. Micro-Doppler detection mainly identifies drones through the micro-Doppler effect generated by a third-party signal source and rotor equipment. However, it is limited by the third-party signal source and the type of drone, resulting in relatively low engineering applicability. Radio frequency signal identification mainly detects and identifies drones by detecting image transmission and flight control signals.

[0004] Radio frequency (RF) signal identification (RFS) is one of the main methods for drone detection and identification. Drones communicate with ground control stations via RF for flight control and navigation, real-time video transmission, and telemetry information transmission. RF drone detection systems can detect drones by monitoring the communication spectrum. Many scholars have conducted research on deep learning detection of drone RF signals. Allahham developed the DroneRF dataset using three commercial drones for validation testing. Mohammad F designed a DNN based on the DroneRF dataset, validating the feasibility of using a drone RF database for detection and identification, but the recognition rate was not high. S. Al-Emadi designed a CNN network and fully connected layers to further improve the detection and identification accuracy of the DroneRF dataset, increasing drone detection and classification accuracy. Yongguang Mo used compressed sensing to preprocess the DroneRF dataset, then designed DNN and CNN networks for identification and classification respectively. This multi-level detection method achieved an accuracy of over 99% in detecting the presence, type, and flight model of drones. These studies were conducted on a limited dataset and did not investigate the impact of noise on detection performance. Furthermore, they did not investigate detection performance in the presence of multiple drone signals or in cases of differences or interference at the drone's radio frequency end after deployment.

[0005] For radio frequency identification (RFID), the RFID fingerprint characteristics will be different due to the model, batch or manufacturing time. The existing technical solutions are based on deep networks trained on a small amount of publicly available experimental data, which have defects such as poor generalization ability, resulting in a decrease in recognition rate. The current technical solutions cannot identify multiple drones, especially the identification of drones of the same model. Summary of the Invention

[0006] The technical problem this invention aims to solve is to overcome the shortcomings of existing technologies and provide a drone RFID method based on a deep attention detection model. Employing a domain adaptation-based approach, this method effectively addresses the generalization capability problem of deep networks and can simultaneously identify multiple drones, particularly those of the same model. To achieve this objective, this invention provides a drone RFID method based on a deep attention detection model, comprising: acquiring multiple drone RFID signals to be identified; inputting the multiple drone RFID signals to be identified into a pre-trained deep attention detection model, and predicting the output RFID signal classification; the RFID signal classification includes the presence of a drone, the drone model, and the current drone operating mode. Preferably, the multiple drone RFID signals to be identified are multiple drone RFID signals of the same model acquired at the same time or drone RFID signals of different batches of the same model acquired at different time periods. Firstly, a deep attention detection model is pre-trained, achieved through the following steps: Constructing the deep attention detection model; obtaining a training set, which includes drone radio frequency (RF) signal samples, the corresponding real drone models, and the corresponding real drone operating modes; preprocessing the drone RF signal samples; using the drone RF signal samples as input to the deep attention detection model, and the corresponding real drone models and operating modes as outputs, thus establishing a mapping relationship between the drone RF signal samples, their corresponding real drone models, and their corresponding real drone operating modes; iteratively updating the network weight parameters of the deep attention detection model using a cross-entropy loss function; if the cross-entropy loss function converges to a certain value, stopping the iterative update, and obtaining the final deep attention detection model.

[0007] Prior to this, the drone radio frequency signal samples include radio frequency signal samples of the same model of drone and radio frequency signal samples of different models of drone;

[0008] Preprocessing includes segmented STFT, spectrum reconstruction, ridge value extraction, noise filtering, and anti-aliasing.

[0009] Priority is that the pre-processed drone radio frequency signal samples are stored in a tree structure to distinguish drones of the same model, multiple batches and time periods;

[0010] The tree structure uses binary unique identifiers to represent drones of the same model, in multiple batches, and at multiple time periods.

[0011] Prior to this, the deep attention detection model employs a fully convolutional network with an encoder-decoder architecture;

[0012] The encoder consists of a Conv convolutional layer, a MaxPool2D layer and a residual network connected in sequence. The Conv convolutional layer has a 3*3 convolutional kernel, the image size of the Conv convolutional layer is 128*128, and the image size of the MaxPool2D layer is 64*64.

[0013] The residual network consists of four layers: ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4. These layers are connected sequentially. The image size of ResNetBlock1 is 64*64, while the image sizes of ResNetBlock2, ResNetBlock3, and ResNetBlock4 are 32*32.

[0014] Preferably, the decoder includes an RF feature module and an RF channel attention aggregation module;

[0015] The radio frequency characteristic module includes FF Block1, FF Block2 and FF Block3 modules. The output of ResnetBlock1 layer is connected to the input of FF Block1 module, the output of ResnetBlock2 layer is connected to the input of FF Block2 module, and the output of ResnetBlock3 layer is connected to the input of FF Block3 module.

[0016] The radio frequency channel attention aggregation module includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an FCA BLOCK1 module, and an FCA BLOCK2 module.

[0017] The output of the FF Block1 module, the first convolutional layer, and the second convolutional layer are connected sequentially. The output of the third convolutional layer is connected to the input of the FF Block1 module, the input of the third convolutional layer is connected to the output of the FCA Block1 module, and the input of the FCA Block1 module is connected to the output of the FF Block2 module.

[0018] The first convolutional layer has an image size of 128*128 and a 3*3 kernel. The second convolutional layer has an image size of 1*128 and a 1*3 kernel. The third and fourth convolutional layers have images of 32*32. The input of the FF Block3 module is connected to the fourth convolutional layer with a 1×1 kernel.

[0019] The input of the fourth convolutional layer is connected to the output of the FCA BLOCK2 module, and the input of the FCA BLOCK2 module is connected to the output of the ResnetBlock4 layer.

[0020] Preferably, ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4 all include an input x, a first Conv convolutional layer, a first ReLU layer, a second Conv convolutional layer, a second ReLU layer, and a third Conv convolutional layer. These layers are sequentially connected, and the third Conv convolutional layer is added to and fused with the input x before being connected to the third ReLU layer. The RF channel attention aggregation module includes an FCABLOCK1 module and an FCA BLOCK2 module. The FCA BLOCK1 module and the FCA... Each BLOCK2 module includes parallel-connected positional relationship attention blocks and channel relationship attention blocks. The positional relationship attention blocks include the fifth, sixth, seventh, and eighth convolutional layers. The output of the fifth convolutional layer is connected to the inputs of the sixth, seventh, and eighth convolutional layers. The outputs of the sixth and seventh convolutional layers are multiplied to obtain the first output; the first output is multiplied by the output of the eighth convolutional layer to obtain the second output; and the second output is added to the output of the fifth convolutional layer to obtain the third output. The channel relationship attention blocks include... The system consists of nine, ten, eleven, twelfth, and thirteenth convolutional layers. The output of the ninth convolutional layer is connected to the inputs of the tenth, eleventh, and twelfth convolutional layers. The outputs of the tenth and eleventh convolutional layers are multiplied to obtain the fourth output. The fourth output is multiplied by the output of the twelfth convolutional layer to obtain the fifth output. The fifth output is input into the thirteenth convolutional layer to obtain the sixth output. The sixth output is added to the output of the ninth convolutional layer to obtain the seventh output. The seventh output is added to the third output to obtain the outputs of the FCA BLOCK1 and FCA BLOCK2 modules.Prioritize the pre-training of a deep attention detection model, achieved through the following steps: Obtain a test set, including UAV RF signal test samples, corresponding real UAV models, and corresponding real UAV operating modes; input the UAV RF signal test samples into the final deep attention detection model to predict and output UAV models and operating modes; count the number of predicted UAV models that match the corresponding real UAV models; calculate the ratio of the number of predicted UAV models that match the corresponding real UAV models to the total number of all UAV models predicted by the final deep attention detection model, obtaining the final deep attention detection model's UAV model accuracy; count the number of predicted UAV operating modes that match the corresponding real UAV operating modes; calculate the ratio of the number of predicted UAV operating modes that match the corresponding real UAV operating modes to the total number of all UAV operating modes predicted by the final deep attention detection model, obtaining the final deep attention detection model's UAV operating mode accuracy; if the UAV model accuracy is higher than a set UAV model accuracy threshold and the UAV operating mode accuracy is higher than a set UAV operating mode accuracy threshold, then the final deep attention detection model is deemed qualified. An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the program, implements the steps of any of the methods described above. A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0021] The beneficial effects achieved by this invention are as follows:

[0022] This invention proposes a method for detecting and identifying UAV radio frequency signals using a software-defined radio device. The software-defined radio device identifies and collects UAV radio frequency signals, and a deep attention detection model is used for identification to achieve the purpose of identifying UAVs.

[0023] This invention proposes a drone radio frequency identification (RFID) method based on a deep attention model. A detection model based on this model is proposed and validated through comprehensive comparison on the publicly available DroneRF dataset. This further verifies the feasibility of simultaneously identifying multiple drones of the same model via RFID, particularly addressing the issue of multi-drone signal detection. Furthermore, based on this method, a rapid RFID device for drones is constructed, capable of detecting the presence of drones and distinguishing the operational status of multiple different drones.

[0024] This invention uses software-defined radio equipment to collect radio frequency signals from two drones of the same model, and a deep attention detection model to identify the working mode, which can distinguish multiple drones of the same model. Attached Figure Description

[0025] Figure 1 This is a block diagram illustrating the principle of the present invention for acquiring radio frequency signals from a drone;

[0026] Figure 2 This is a distribution diagram of the tree-structured storage of the UAV in this invention;

[0027] Figure 3 This is an architecture diagram of the encoder-decoder architecture used in this invention;

[0028] Figure 4 This is an architecture diagram of the residual network module in this invention;

[0029] Figure 5 This is an architecture diagram of the radio frequency channel attention aggregation module in this invention. Detailed Implementation

[0030] The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0031] Example 1

[0032] There are many models of edge computing devices and software radio devices available in the prior art. Those skilled in the art can select the appropriate model according to actual needs. This embodiment will not list them all.

[0033] Example 2

[0034] The drone's radio frequency (RF) signals reflect its status, including whether it is powered off, powered on, connected, hovering, flying, and displaying video. These RF signals are received by an RF receiver, analyzed, and the data is collected as follows: Figure 1 As shown.

[0035] A novel radio frequency identification (RFID) method for unmanned aerial vehicles (UAVs) based on a deep attention model is proposed. This method employs a fully convolutional network with an encoder-decoder architecture, using a residual network as the backbone for feature extraction. A specially designed attention aggregation module for the radio frequency channel is incorporated for identification. The model is trained and validated on the publicly available DroneRF dataset, achieving 100% accuracy in UAV detection and classification, 99.695% in UAV recognition, and 99.633% in UAV operating mode recognition. The network weight parameters of the deep attention detection model include the network training parameters w and bias parameters b, as well as α and β from the global target loss parameters.

[0036] This invention uses the software-defined radio device USRP_X310 to collect radio frequency data from two drones of the same model. After transfer learning, it can detect drones and identify their operating modes.

[0037] Software-defined radio (SDR) devices typically use USRP RF receivers, such as the USRP_X310, to collect UAV RF signals. The SDR devices then preprocess the UAV RF signals, including segmented STFT, spectrum reconstruction, ridge value extraction, noise filtering, and anti-aliasing, to facilitate the digital storage of the UAV RF signals.

[0038] The preprocessed UAV radio frequency signals are stored in a specific format, distinct from publicly available data structures. The storage format for the preprocessed UAV radio frequency signals in this invention employs a five-level tree structure, such as... Figure 2 As shown, drones of the same model, from multiple batches, and from multiple time periods are distinguished. The pre-processed drone RF signals are organized and stored in a tree structure, and are marked with a binary unique identifier (BUI).

[0039] The storage is organized in a tree structure with 5 levels: Level 1, Level 2, Level 3, Level 4, and Level 5.

[0040] Specifically, the first field of the BUI represents the time period, the second field represents the category, the third field represents the batch, the fourth field represents the working mode, and the fifth field or other extensions represent the distance.

[0041] Level 1 includes Day 1, ..., Day N. The binary unique identifier BUI of Day 1 is 0-XXXX, and the binary unique identifier BUI of Day N is nXXXX.

[0042] Level 2 includes Category 1, ..., Category N. The binary unique identifier BUI of Category 1 is 0-1-XXX, and the binary unique identifier BUI of Category N is 0-nXXX.

[0043] Level 3 includes UAV1, UAV2, ..., UAVN. The binary unique identifier of UAV1 is BUI = 0-1-0-XX, the binary unique identifier of UAV2 is BUI = 0-1-1-XX, and the binary unique identifier of UAVN is BUI = 0-1-nXX.

[0044] Level 4 includes Model 1, Model 2, Model 3 and Model 4. The binary unique identifier BUI of Model 1 is 0-1-1-0-X, the binary unique identifier BUI of Model 2 is 0-1-1-1-X, the binary unique identifier BUI of Model 3 is 0-1-1-2-X, and the binary unique identifier BUI of Model 4 is 0-1-1-nX.

[0045] In Level 5, the binary unique identifier BUI = 01100 for Distance 1 corresponding to Model 1, the binary unique identifier BUI = 01101 for Distance 2 corresponding to Model 1, ..., the binary unique identifier BUI = 0110m for Distance m corresponding to Model 1;

[0046] In Level 5, the binary unique identifier BUI = 01110 for Distance 1 corresponding to Model 2, the binary unique identifier BUI = 01111 for Distance 2 corresponding to Model 2, ..., the binary unique identifier BUI = 0111m for Distance m corresponding to Model 2;

[0047] In Level 5, the binary unique identifier BUI = 01120 for Distance 1 corresponding to Model 3, the binary unique identifier BUI = 01121 for Distance 2 corresponding to Model 3, ..., the binary unique identifier BUI = 0112m for Distance m corresponding to Model 3;

[0048] In Level 5, the binary unique identifier BUI = 011n0 for Distance 1 corresponding to Model 4, the binary unique identifier BUI = 011n1 for Distance 2 corresponding to Model 4, ..., the binary unique identifier BUI = 011nm for Distance m corresponding to Model 4.

[0049] Fully convolutional networks based on deep attention models using an encoder-decoder architecture, such as Figure 3 As shown, it mainly consists of a feature extraction path (encoder) and an upsampling path (decoder).

[0050] Fully convolutional networks based on deep attention models using an encoder-decoder architecture, such as Figure 3 As shown, it mainly consists of a feature extraction path (encoder) and an upsampling path (decoder).

[0051] The encoding path includes a Conv convolutional layer, a MaxPool2D layer, and a residual network. These layers are connected sequentially. The Conv convolutional layer uses a 3x3 kernel. The image size for the Conv convolutional layer is 128x128, and the image size for the MaxPool2D layer is 64x64.

[0052] In the encoding path, a residual network (ResNet101) is used as the backbone for feature extraction. Without changing the scale of the pre-trained parameters, the focus is on preserving more detailed information of the UAV radio frequency signal. The downsampling layers in the last two residual blocks are replaced by dilated convolutional layers. This strategy can preserve more category information.

[0053] Specifically, the residual network includes ResnetBlock1, ResnetBlock2, ResnetBlock3 and ResnetBlock4 layers. The image size of ResnetBlock1 is 64*64, and the image sizes of ResnetBlock2, ResnetBlock3 and ResnetBlock4 are 32*32.

[0054] The decoding path mainly consists of a radio frequency feature (FF BLOCK) module and a radio frequency channel attention aggregation (FCA) module connected in a specific order. First, corresponding to the residual blocks of ResNet101, the FF BLOCK module includes three FF modules: FF Block1, FF Block2, and FF Block3. The image size of FF Block1 is 64*64, while the image sizes of FF Block2 and FF Block3 are 32*32. The input of FF Block1 is connected to the output of the ResNetBlock1 layer, the input of FF Block2 is connected to the output of the ResNetBlock2 layer, and the input of FF Block3 is connected to the output of the ResNetBlock3 layer. This process improves the recognition capability of radio frequency detail information by generating global feature guidance.

[0055] The RF channel attention aggregation module is embedded in the header and middle of the decoding path to capture long-range background information in the spatial and channel domains, respectively. The header of the FCA module includes a first convolutional layer, a second convolutional layer, a third convolutional layer, an FCABLOCK1 module, a fourth convolutional layer, and an FCA BLOCK2 module. The FF Block1 module, the first convolutional layer, and the second convolutional layer are connected sequentially. The image size of the first convolutional layer is 128*128, and the convolutional kernel of the first convolutional layer is 3*3. The image size of the second convolutional layer is 1*128, and the convolutional kernel of the second convolutional layer is 1*3.

[0056] In addition, the input layer of the FF Block1 module is connected to a third convolutional layer with a 1×1 kernel, and the input of the third convolutional layer is connected to a fourth convolutional layer with a 1×1 kernel. This is used for channel dimension transformation to match low-level feature maps. The image size of the third and fourth convolutional layers is 32*32.

[0057] Finally, after the semantic feature map is output from the FF Block1 module, it is sequentially fed into the first and second convolutional layers for bilinear upsampling to obtain a fully connected classification map. It's important to note that normalization and ReLU activation functions are applied after each convolutional layer and its transposed form.

[0058] ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4 layers are used as feature extraction paths to learn effective categorical features. As the unit structure of the residual network, ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4 layers represent a mapping process of several stacked convolutional layers, such as... Figure 2 As shown, each includes a first Conv convolutional layer, a first ReLU layer, a second Conv convolutional layer, a second ReLU layer, and a third Conv convolutional layer. The first Conv convolutional layer, the first ReLU layer, the second Conv convolutional layer, the second ReLU layer, and the third Conv convolutional layer are connected in sequence. The output of the third Conv convolutional layer and the input x (radio frequency signal matrix) feature are fused and then passed through the third ReLU layer to obtain H(x) (the radio frequency signal matrix after convolution).

[0059] The attention mechanism enables the method of this invention to focus attention on key regions related to specific categories, enhancing discriminative features by encoding background information, which is an effective way to improve the performance of radio frequency feature classification. Therefore, the deep attention detection model in this invention proposes a radio frequency channel attention aggregation module (FCA), such as... Figure 5 As shown, this module consists of parallel-connected position relation attention locks and channel relation attention locks.

[0060] Figure 5 The convolutions in the diagram use both 1x1 and 3x3 convolutions interchangeably, and the arrows indicate the changes in convolution size and the ReLU activation function.

[0061] Gradient descent algorithms progressively search and set parameters by reducing the loss function. This invention utilizes multiple cross-entropy loss functions to monitor features at specific scales in the network. This deep monitoring strategy can distinguish multi-scale features to capture specific category context and optimize the training process.

[0062] The cross-entropy loss function represents the deviation between the predicted value and the true label value at each measurement point, and the formula is as follows:

[0063]

[0064] L represents the deviation between the predicted value and the true label value of the deep attention detection model. This represents the actual label value (the fourth field of the aforementioned BUI). This represents the predicted value, where N is the convolution length, for example, 32*32=1024;

[0065] A weighted model was established, and its global objective loss function is expressed as follows, where α and β represent weight coefficients. During initial training, α and β were set to 0.4 and 0.2, respectively. Lp1, Lp2, and Lo are respectively... Figure 3 If we have three cross-entropy loss function values, then the global objective loss function is L.

[0066] L=lo+αLp1+βLp2(2)

[0067] Model training and testing:

[0068] 1. Dataset partitioning

[0069] The initial lab dataset was divided into three categories: drone detection, drone recognition, and drone pattern recognition. Using 10x cross-validation, the dataset was randomly divided into 10 subsets: 9 subsets for training and 1 subset for testing.

[0070] 2. Training Framework and Parameters. The training framework uses deep learning frameworks such as PyTorch, TensorFlow, MXNet, or PaddlePaddle. Training is performed on a computer (not limited to) equipped with an Intel Core i7-9700k CPU, 32GB RAM, and an NVIDIA RTX3080 GPU (12GB VRAM). Training parameters are shown in Table 1, and the training optimizer is Adam. The training run consists of 100 epochs.

[0071] Table 1 Parameter Settings

[0072] Experimental categories Training rounds Batch size Output function loss function Initial learning rate Drone Inspection 30 10 Sigmoid mse 0.001 Drone identification 200 25 Softmax CrossEntropy 0.00003 Working mode recognition 300 50 Softmax CrossEntropy 0.00003

[0073] 3. Testing

[0074] Test the model using the test set. If the recognition rate is too low, adjust the model parameters and retrain. Once 95% is achieved, stop training and save the model.

[0075] 4. Model Iterative Improvement

[0076] When new drone models or other test data are available, transfer learning is used to continue training on the previously saved model, allowing for rapid model iteration.

[0077] The training model is integrated into an edge computing device, combined with a device for detecting the radio frequency signals of drones, to perform radio frequency signal identification of drones.

[0078] 1. This invention proposes a method for detecting and identifying UAV radio frequency signals by combining software radio equipment. The software radio equipment collects UAV signals and uses a deep attention model for identification to achieve the purpose of identifying UAVs.

[0079] 2. This invention proposes a drone radio frequency identification (RFID) method based on a deep attention model. It employs a fully convolutional network based on an encoder-decoder architecture, uses a residual network as the backbone network for feature extraction, and specifically designs a radio frequency channel attention aggregation (FCA) module for identification. The model is trained and validated on a public dataset (DroneRF dataset), achieving a drone detection and classification rate of 99.895%, a drone recognition rate of 98.33%, and a drone operating mode recognition rate of 99.33%.

[0080] 3. The method of this invention collects radio frequency data from two drones of the same model. After transfer learning, it can detect drones and identify their working modes.

[0081] 4. This invention utilizes edge computing devices to deploy trained models for drone recognition, thereby expanding application scenarios.

[0082] 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, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] 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.

[0084] 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.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for radio frequency identification (RFID) of unmanned aerial vehicles (UAVs) based on a deep attention detection model, characterized in that, include: Collect radio frequency signals from multiple drones to be identified; Multiple unmanned aerial vehicle (UAV) radio frequency (RF) signals to be identified are input into a pre-trained deep attention detection model to predict the classification of the output RF signals. Radio frequency signal classification includes the presence of a drone, the drone model, and the current drone operating mode; The deep attention detection model employs a fully convolutional network with an encoder-decoder architecture, wherein the decoder includes a radio frequency feature module and a radio frequency channel attention aggregation module. The radio frequency characteristic module includes FF Block1 module, FF Block2 module and FF Block3 module; The output of the ResnetBlock1 layer is connected to the input of the FF Block1 module, the output of the ResnetBlock2 layer is connected to the input of the FF Block2 module, and the output of the ResnetBlock3 layer is connected to the input of the FF Block3 module. The radio frequency channel attention aggregation module includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, an FCABLOCK1 module, and an FCA BLOCK2 module; The output of the FF Block1 module, the first convolutional layer, and the second convolutional layer are connected in sequence. The output of the third convolutional layer is connected to the input of the FF Block1 module. The input of the third convolutional layer is connected to the output of the FCA Block1 module. The input of the FCA Block1 module is connected to the output of the FF Block2 module. The image size of the first convolutional layer is 128*128, and the kernel size of the first convolutional layer is 3*3. The image size of the second convolutional layer is 1*128, and the kernel size of the second convolutional layer is 1*3. The image size of the third and fourth convolutional layers is 32*32. The input of the FF Block3 module is connected to the fourth convolutional layer with a 1×1 kernel; The input of the fourth convolutional layer is connected to the output of the FCA BLOCK2 module, and the input of the FCA BLOCK2 module is connected to the output of the ResnetBlock4 layer.

2. The UAV radio frequency identification method based on a deep attention detection model according to claim 1, characterized in that, Multiple unidentified drone radio frequency signals can be either multiple drone radio frequency signals of the same model acquired at the same time, or drone radio frequency signals of the same model from different batches acquired at different time periods.

3. The UAV radio frequency identification method based on a deep attention detection model according to claim 1, characterized in that, A deep attention detection model is pre-trained through the following steps: Build a deep attention detection model; Obtain the training set, which includes drone radio frequency signal samples, the real drone models corresponding to the drone radio frequency signal samples, and the real drone operating modes corresponding to the drone radio frequency signal samples. Preprocessing of UAV radio frequency signal samples; The drone radio frequency signal samples are used as input to the deep attention detection model, and the real drone models and working modes corresponding to the drone radio frequency signal samples are used as outputs of the deep attention detection model. The mapping relationship between the drone radio frequency signal samples, the real drone models and working modes corresponding to the drone radio frequency signal samples is constructed. The network weight parameters of the deep attention detection model are iteratively updated using the cross-entropy loss function. If the cross-entropy loss function converges to a certain value, the iterative update stops, and the final deep attention detection model is obtained.

4. The UAV radio frequency identification method based on a deep attention detection model according to claim 3, characterized in that, The drone radio frequency signal samples include radio frequency signal samples of the same model of drone and radio frequency signal samples of different models of drone; Preprocessing includes segmented STFT, spectrum reconstruction, ridge value extraction, noise filtering, and anti-aliasing.

5. The UAV radio frequency identification method based on a deep attention detection model according to claim 4, characterized in that, The preprocessed drone radio frequency signal samples are stored in a tree structure to distinguish drones of the same model, multiple batches, and multiple time periods; The tree structure uses binary unique identifiers to represent drones of the same model, in multiple batches, and at multiple time periods.

6. The UAV radio frequency identification method based on a deep attention detection model according to claim 3, characterized in that, The deep attention detection model employs a fully convolutional network with an encoder-decoder architecture; The encoder consists of a Conv convolutional layer, a MaxPool2D layer and a residual network connected in sequence. The Conv convolutional layer has a 3*3 convolutional kernel, the image size of the Conv convolutional layer is 128*128, and the image size of the MaxPool2D layer is 64*64. The residual network consists of four layers: ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4. These layers are connected sequentially. The image size of ResNetBlock1 is 64*64, while the image sizes of ResNetBlock2, ResNetBlock3, and ResNetBlock4 are 32*32.

7. The UAV radio frequency identification method based on a deep attention detection model according to claim 1, characterized in that, ResNetBlock1, ResNetBlock2, ResNetBlock3, and ResNetBlock4 all include an input x, a first Conv convolutional layer, a first ReLU layer, a second Conv convolutional layer, a second ReLU layer, and a third Conv convolutional layer. The first Conv convolutional layer, the first ReLU layer, the second Conv convolutional layer, the second ReLU layer, and the third Conv convolutional layer are connected sequentially. The third Conv convolutional layer and the input x are added together and fused to the third ReLU layer. The RF channel attention aggregation module includes FCA BLOCK1 and FCA BLOCK2 modules. Both FCA BLOCK1 and FCA BLOCK2 modules include parallel-connected positional relationship attention blocks and channel relationship attention blocks. The positional relationship attention block includes the fifth, sixth, seventh, and eighth convolutional layers. The output of the fifth convolutional layer is connected to the input of the sixth, seventh, and eighth convolutional layers. The output of the sixth and seventh convolutional layers are multiplied to obtain the first output. The first output is multiplied by the output of the eighth convolutional layer to obtain the second output. The second output is added to the output of the fifth convolutional layer to obtain the third output. The channel relationship attention block includes the ninth, tenth, eleventh, twelfth and thirteenth convolutional layers. The output of the ninth convolutional layer is connected to the input of the tenth, eleventh and twelfth convolutional layers. The output of the tenth and eleventh convolutional layers are multiplied to obtain the fourth output. The fourth output is multiplied by the output of the twelfth convolutional layer to obtain the fifth output; The fifth output is input into the thirteenth convolutional layer to obtain the sixth output; The sixth output is added to the output of the ninth convolutional layer to obtain the seventh output. The seventh output and the third output are added together to obtain the outputs of the FCA BLOCK1 module and the FCA BLOCK2 module.

8. The UAV radio frequency identification method based on a deep attention detection model according to claim 2, characterized in that, A deep attention detection model is pre-trained through the following steps: Obtain the test set, which includes drone radio frequency signal test samples, the real drone models corresponding to the drone radio frequency signal test samples, and the real drone operating modes corresponding to the drone radio frequency signal test samples. The test samples of the drone's radio frequency signal are input into the final deep attention detection model to predict and output the drone model and drone operating mode. Count the number of predicted drone models that are the same as the corresponding real drone models; The ratio of the number of predicted drone models that are the same as the corresponding real drone models to the total number of all drone models predicted by the final deep attention detection model is used to obtain the drone model accuracy of the final deep attention detection model. Count the number of predicted drone operating modes that are the same as the corresponding real drone operating modes; The ratio of the number of predicted drone operating modes that are the same as the corresponding real drone operating modes to the total number of all drone operating modes predicted by the final deep attention detection model is used to obtain the drone operating mode accuracy of the final deep attention detection model. If the accuracy rate of the drone model is higher than the set threshold for drone model accuracy and the accuracy rate of the drone working mode is higher than the set threshold for drone working mode accuracy, then the final deep attention detection model is deemed qualified.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images