Unmanned aerial vehicle radio frequency fingerprint identification method and device, medium and equipment
By combining time-frequency anisotropic convolution and multi-scale pyramid structure with a meta-learning framework, the problem of low recognition accuracy of UAV radio frequency fingerprint recognition in complex electromagnetic environments is solved, and efficient UAV model recognition is achieved.
Patent Information
- Application Number
- CN202511489450.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-01-13
AI Technical Summary
Existing drone radio frequency fingerprinting technology has low recognition accuracy in complex electromagnetic environments. Deep learning models rely on large-scale datasets, resulting in insufficient training samples. Meta-learning methods fail to effectively capture the inherent coupling relationship between temporal dynamic characteristics and spectral distortion caused by hardware defects, leading to low recognition accuracy.
By employing time-frequency anisotropic convolution combined with a multi-scale pyramid structure and a meta-learning framework, time-domain and frequency-domain features are extracted using convolution kernels of different sizes. Combined with channel attention mechanism and model-independent meta-learning, efficient extraction and recognition of UAV radio frequency fingerprint features are achieved.
It improves the accuracy and robustness of UAV radio frequency fingerprint recognition, enabling it to quickly adapt to new devices and dynamic environments under limited sample conditions, reducing model complexity and improving recognition accuracy.
Smart Images

Figure CN121327485A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer, in particular to a UAV radio frequency fingerprint identification method, device, medium and equipment. BACKGROUND
[0002] At present, with the popularity of UAVs in aerial photography, agricultural monitoring, geographic mapping and other fields, the security risks such as airspace intrusion, privacy invasion and illegal load delivery caused by UAVs are increasingly prominent. The existing UAV identification methods mainly include acoustic feature analysis, visual recognition and Automatic Dependent Surveillance-Broadcast (ADS-B) system, but the acoustic method is easily disturbed by environmental noise and has limited action distance, the visual recognition is limited in performance under bad weather conditions, and the ADS-B system has the risk of fraud and is not popular in small UAVs. Compared with the above, radio frequency fingerprint identification has the advantages of non-contact detection, long-distance operation and dynamic environment adaptability, and is especially suitable for identity authentication of encrypted signals or non-cooperative UAVs. As a non-intrusive identity authentication method, UAV radio frequency fingerprint identification technology has received widespread attention in the security field in recent years. This technology relies on the radio frequency fingerprint features generated by the inherent defects of the hardware circuit in the UAV communication link, and realizes device differentiation by analyzing the unintentional modulation differences in the radio frequency signal.
[0003] In the field of radio frequency fingerprint identification in the prior art, traditional machine learning methods rely on manually designed time-frequency domain features (such as energy transient characteristics, instantaneous amplitude and phase features, power spectral density based on short-time Fourier transform), first extract the time-frequency features of the UAV radio frequency signal, and then realize identification combined with support vector machine, random forest and other classifiers. This kind of method performs well in static environment, but is sensitive to channel noise in complex electromagnetic environment, and the recognition accuracy decreases significantly.
[0004] Deep learning technology automatically learns feature representation through convolutional neural network, such as a research scheme that uses adaptive attention mechanism to enhance the extraction ability of key time-frequency regions. Under sufficient data conditions, it can achieve high accuracy. However, deep models rely heavily on large-scale high-quality data sets, and in actual scenarios, the cost of collecting UAV radio frequency signals is high, resulting in insufficient training samples, unbalanced class distribution, and causing model overfitting and weakening of generalization ability.
[0005] To solve the problem of data scarcity, few-shot learning schemes have emerged. Meta-learning methods such as MLGPN achieve few-shot open-set identification through metric learning, but their single-domain (time domain or frequency domain) signal processing mechanism cannot effectively capture the internal coupling relationship between time-domain dynamic characteristics and spectral distortion caused by hardware defects, resulting in low accuracy of UAV radio frequency fingerprint identification. SUMMARY
[0006] Therefore, it is necessary to provide a method, device, medium and equipment for unmanned aerial vehicle radio frequency fingerprint identification in view of the above technical problems.
[0007] The application adopts the following technical solutions: The application provides a method for unmanned aerial vehicle radio frequency fingerprint identification, which comprises the following steps: acquiring a radio frequency signal of an unmanned aerial vehicle and determining a time-frequency graph corresponding to the radio frequency signal; inputting the time-frequency graph into a radio frequency fingerprint identification network; performing time domain convolution along a time axis of the time-frequency graph through a first convolution kernel of the radio frequency fingerprint identification network, extracting time domain features of the time-frequency graph, performing frequency domain convolution along a frequency axis of the time-frequency graph through a second convolution kernel of the radio frequency fingerprint identification network, extracting frequency domain features of the time-frequency graph, and completing time-frequency anisotropic convolution; the first convolution kernel is different in size from the second convolution kernel; finally, splicing the time domain features and the frequency domain features to obtain spliced features, extracting radio frequency fingerprint features of the unmanned aerial vehicle through multi-scale time-frequency anisotropic convolution in the radio frequency fingerprint identification network according to the spliced features, and determining a model of the unmanned aerial vehicle according to the radio frequency fingerprint features.
[0008] The application provides a device for unmanned aerial vehicle radio frequency fingerprint identification, which comprises the following steps: An acquisition module is configured to acquire a radio frequency signal of an unmanned aerial vehicle and determine a time-frequency graph corresponding to the radio frequency signal; A convolution module is configured to input the time-frequency graph into a radio frequency fingerprint identification network, perform time domain convolution along a time axis of the time-frequency graph through a first convolution kernel of the radio frequency fingerprint identification network, extract time domain features of the time-frequency graph, perform frequency domain convolution along a frequency axis of the time-frequency graph through a second convolution kernel of the radio frequency fingerprint identification network, extract frequency domain features of the time-frequency graph, and complete time-frequency anisotropic convolution; the first convolution kernel is different in size from the second convolution kernel; An identification module is configured to splice the time domain features and the frequency domain features to obtain spliced features, extract radio frequency fingerprint features of the unmanned aerial vehicle through multi-scale time-frequency anisotropic convolution in the radio frequency fingerprint identification network according to the spliced features, and determine a model of the unmanned aerial vehicle according to the radio frequency fingerprint features.
[0009] The application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the above-mentioned method for unmanned aerial vehicle radio frequency fingerprint identification.
[0010] The application provides a computer device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the above-mentioned method for unmanned aerial vehicle radio frequency fingerprint identification when executing the program.
[0011] The above-mentioned at least one technical solution adopted by the application can achieve the following beneficial effects: The application extracts the time domain features of the time-frequency diagram by sliding along the time axis of the time-frequency diagram for time domain convolution, and extracts the frequency domain features of the time-frequency diagram by sliding along the frequency axis of the time-frequency diagram for frequency domain convolution, and the time domain convolution and the frequency domain convolution adopt heterogeneous combination convolution kernels of different sizes to realize time-frequency anisotropic convolution, respectively capture the short-time dynamic evolution and local energy distribution characteristics of the radio frequency signal of the unmanned aerial vehicle, avoid the uniform operation defects of the traditional isotropic convolution, decouple through the physically interpretable dimension, reduce the parameter redundancy while enhancing the feature discriminability, and improve the unmanned aerial vehicle radio frequency fingerprint identification accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0012] The accompanying drawings, which are included to provide a further understanding of the application, illustrate embodiments of the application and together with the description explain the application. The accompanying drawings illustrate embodiments of the application and together with the description explain the application. In the drawings:
[0013] Figure 1 A flowchart of an unmanned aerial vehicle radio frequency fingerprint identification method provided by the application is shown in the figure; Figure 2 A schematic diagram of a lightweight time-frequency diagram classification network structure based on time-frequency anisotropic convolution and channel attention mechanism provided by the application is shown in the figure; Figure 3 A time-frequency anisotropic convolution block module diagram provided by the application is shown in the figure; Figure 4 A model offline training and small sample self-adaptive test mode provided by the application is shown in the figure; Figure 5 A schematic diagram of an unmanned aerial vehicle radio frequency fingerprint identification device provided by the application is shown in the figure. DETAILED DESCRIPTION
[0014] To make the purpose, technical scheme and advantages of the application clearer, the technical scheme of the application will be described below in conjunction with specific embodiments of the application and corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.
[0015] Currently, to solve the problem of data scarcity, few-shot learning schemes are gradually emerging. Typical representatives include: unsupervised pre-training methods based on mask autoencoders (MAE) alleviate overfitting by reconstructing signals, but still require a certain amount of data to support; triple residual twin network (TRSN) uses residual blocks to construct a feature mapping network to enhance discriminability, but the model complexity is high; light-weight schemes based on triple loss convolutional network (TLCNN) combine unsupervised clustering to reduce complexity, but are not robust in noisy scenarios. Recent meta-learning methods such as MLGPN achieve few-shot open-set recognition through metric learning, but their single-domain (time domain or frequency domain) signal processing mechanism fails to effectively capture the intrinsic coupling relationship between time-domain dynamic characteristics and spectral distortion caused by hardware defects, limiting the improvement of feature discriminability.
[0016] That is, existing unmanned aerial vehicle radio frequency fingerprint identification technologies have significant limitations. First, data augmentation-based methods (such as Mixup linear transformation) cannot effectively simulate the electromagnetic environment characteristics of real complex scenarios, which can easily distort sample features and lose key fingerprint information, and the generalization ability in few-shot scenarios is limited. Second, schemes that rely on complex deep models (such as ARFNet adaptive networks) have a sharp decline in recognition performance when training data is insufficient, especially in few-shot conditions, and the model generalization ability is severely insufficient, making it difficult to adapt to new devices or dynamic environments. Third, existing spectral filtering techniques can suppress certain noise, but their effectiveness significantly decreases in low signal-to-noise ratio or complex electromagnetic interference environments, making it difficult to ensure the robustness of fingerprint features. In addition, advanced meta-learning methods (such as MLGPN) only process single-domain (time domain or frequency domain) signals and fail to capture the intrinsic coupling relationship between time-domain dynamic characteristics and spectral distortion caused by hardware defects, limiting the discriminability of features.
[0017] The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.
[0018] Figure 1 The present application is a flowchart of an unmanned aerial vehicle radio frequency fingerprint identification method, which specifically includes the following steps: S101: Obtain the radio frequency signal of the unmanned aerial vehicle and determine the time-frequency graph corresponding to the radio frequency signal.
[0019] S102: Input the time-frequency graph into the radio frequency fingerprint identification network, perform time-domain convolution along the time axis of the time-frequency graph by sliding the first convolution kernel of the radio frequency fingerprint identification network, extract the time-domain features of the time-frequency graph, perform frequency-domain convolution along the frequency axis of the time-frequency graph by sliding the second convolution kernel of the radio frequency fingerprint identification network, extract the frequency-domain features of the time-frequency graph, and complete the time-frequency anisotropic convolution; the size of the first convolution kernel is different from that of the second convolution kernel.
[0020] S103: The time domain features and the frequency domain features are spliced to obtain spliced features, and a radio frequency fingerprint feature of the unmanned aerial vehicle is extracted through multi-scale time-frequency anisotropic convolution in a radio frequency fingerprint identification network according to the spliced features, so as to determine the unmanned aerial vehicle model according to the radio frequency fingerprint feature.
[0021] For the convenience of description, the following will be described taking the server as the execution subject. The server mentioned in the present application can be a server arranged in a service platform, or a device such as a desktop computer, a notebook computer, etc. capable of executing the scheme of the present application.
[0022] The present application proposes a kind of unmanned aerial vehicle radio frequency fingerprint identification method for fusing time-frequency feature decoupling architecture and meta-learning optimization, including signal preprocessing, lightweight network design and meta-learning training mechanism three technical implementations.
[0023] The server can first acquire the radio frequency signal of the unmanned aerial vehicle, and determine the corresponding time-frequency diagram based on the radio frequency signal as the input of the subsequent radio frequency fingerprint identification network.
[0024] Signal preprocessing technical scheme: The server can first convert the original unmanned aerial vehicle in-phase quadrature (I / Q) signal into a time-frequency diagram by short-time Fourier transform (STFT). Specifically, in one or more embodiments of the present application, the radio frequency signal can be processed by framing with a Hanning window, and short-time Fourier transform is performed frame by frame to obtain the frequency spectrum of the radio frequency signal. Then, the modulus value of the frequency spectrum of the radio frequency signal is determined and spliced along the time axis to obtain the time-frequency diagram corresponding to the radio frequency signal.
[0025] For example, the specific implementation can be: the discrete radio frequency signal x [ n ] is processed by framing with a Hanning window , the window length N=2048, and the frame shift is 100% of the window length (no overlap). Short-time Fourier transform is performed frame by frame: .
[0026] Wherein, is the frequency spectrum of the radio frequency signal output by short-time Fourier transform, representing the result in the time-frequency domain; is the discrete time index of the radio frequency signal sequence, taking values 0,1,…,N−1; is the frame index, which determines which frame of radio frequency signal is currently processed; is the discrete frequency index, taking values corresponding to different frequency points, and the value range is 0,1,…,N−1; is the imaginary unit.
[0027] The spectral magnitude can then be calculated and stitched along the time axis to generate a two-dimensional time-frequency graph (time on the horizontal axis, frequency on the vertical axis, and pixel intensity representing energy), outputting a grayscale image of size 224×224 as network input.
[0028] Then, the time-frequency graph can be input into the radio frequency fingerprinting network for radio frequency fingerprint feature extraction. The architecture of this radio frequency fingerprinting network (LightWeightTFNet) is as follows: Figure 2 This diagram illustrates a lightweight time-frequency graph classification network structure based on time-frequency anisotropic convolution and channel attention mechanisms, as described in this invention. The network input is a time-frequency graph, which is then processed through convolutional layers and undergoes three stages of convolution. Stage 1 includes ordinary convolution and a single time-frequency anisotropic convolution. Stages 2 and 3 both include downsampling operations (halving the resolution and doubling the channels) and time-frequency anisotropic convolution. After parallel time-domain / frequency-domain convolution within a block, the output is weighted by channel attention and finally classified through a fully connected layer.
[0029] Time-frequency anisotropic convolution block (TFAnisotropic Block): Use a first convolution kernel of size 5×1 Perform temporal convolution by sliding along the time axis to capture transient fluctuation features. This enables the extraction of time-domain features from the spectrogram. Use a second convolution kernel of size 1×7 Frequency domain convolution is performed by sliding along the frequency axis to extract steady-state spectral distribution features. This enables the extraction of frequency domain features from the spectrogram. Will and After concatenation of the feature space dimension channels, the input is given to the channel attention module to achieve feature fusion. Figure 3 This is a schematic diagram of a time-frequency anisotropic convolutional block module in this invention. Figure 3 middle X Represents the input features, ReLU represents the ReLU activation function, AvgPool represents average pooling, and Sigmoid represents the Sigmoid activation function. X o This indicates the output features.
[0030] Channel attention mechanism: For input features Perform Global Average Pooling (GAP), and then learn the channel weights through the bottleneck structure: .
[0031] in, Dimensional reduction Restoring dimensions, sigmoid activation, and then output the weighted features denotes concatenation, is a channel-wise product.
[0032] Multi-scale pyramid structure: The LightWeightTFNet architecture builds a multi-scale pyramid through three mechanisms to perform multi-scale time-frequency anisotropic convolution to extract the radio frequency fingerprint features of the UAV: cross-layer residual connection, hierarchical down-sampling, and channel dimension expansion. Through three stages, the spatial resolution is gradually reduced, while the channel capacity is expanded, to realize hierarchical feature extraction. The hierarchical down-sampling strategy gradually reduces the spatial resolution through three stages (when the input resolution is 224x224, 224x224→112x112→56x56), as shown in Table 1; the channel expansion strategy synchronously increases the number of channels (32→64→128); finally, the cross-layer residual connection is to fuse the shallow detailed features after down-sampling with the deep abstract features:
[0033] .
[0034] Table 1 Network specific configuration After extracting the radio frequency fingerprint features of the UAV through the multi-scale pyramid structure, classification can be performed through the fully connected layer to determine the UAV model according to the radio frequency fingerprint features.
[0035] Further, in one or more embodiments of the present application, in order to improve the generalization ability of the radio frequency fingerprint recognition network in the few-sample scenario, when training the radio frequency fingerprint recognition network, a model-agnostic meta learning (MAML) framework can be used to train the radio frequency fingerprint recognition network to achieve fast adaptation in the few-sample scenario.
[0036] Specifically, the server can obtain radio frequency signals of the unmanned aerial vehicles as training samples, take each unmanned aerial vehicle model identification as an identification task, and randomly sample a preset number of target identification tasks from a plurality of identification tasks corresponding to the training samples; divide the training samples corresponding to each target identification task into a support set and a query set; then for each target identification task, determine a support set loss of the radio frequency fingerprint identification network through the support set of the target identification task, and perform single-step gradient update on the radio frequency fingerprint identification network to obtain task parameters corresponding to the target identification task; and based on the task parameters corresponding to the target identification task, determine a query set loss of the radio frequency fingerprint identification network through the query set of the target identification task, and train the radio frequency fingerprint identification network with the optimization target of minimizing the query set loss of the target identification task.
[0037] For example, Figure 4 For a model offline training and small sample adaptive testing mode in the application, Figure 4 In represents the inner layer fast adaptation to task i, represents the outer layer meta-optimization to task i, represents the task parameters corresponding to task i. In the training phase, the random stratified sampling under the meta-learning mode is used to construct the task, and the n-way k-shot task auxiliary model is used for adaptive training. In the testing phase, the traditional machine learning mode is used to extract samples for learning and testing, and a multi-task evaluation model is constructed. The following processes can be included:
[0038] (a) Task construction.
[0039] From the training set Random sampling K classification example tasks , i.e. 8 types of unmanned aerial vehicle model identification tasks, each of which includes K sample examples, wherein each task contains a support set (N classes x K samples) and a query set (N classes x Q samples), is the task sampling distribution corresponding to the training sample set D.
[0040] (b) Inner layer fast adaptation.
[0041] For task , calculate the support set loss . Single-step gradient update obtains task parameters: . In the formula, is the original parameter of the radio frequency fingerprint identification network. represents the radio frequency fingerprint identification network based on the parameter .
[0042] (c) Outer layer meta-optimization.
[0043] with Evaluate query set loss . Gradient average update meta-parameter: .
[0044] Based on Figure 1 The unmanned aerial vehicle radio frequency fingerprint identification method shown in the figure, according to the time-frequency diagram corresponding to the radio frequency signal, time domain convolution is carried out along the time axis of the time-frequency diagram, the time domain features of the time-frequency diagram are extracted, and frequency domain convolution is carried out along the frequency axis of the time-frequency diagram, the frequency domain features of the time-frequency diagram are extracted, and the time domain convolution and the frequency domain convolution adopt heterogeneous combination convolution kernels with different sizes to realize time-frequency anisotropic convolution, and respectively capture short-time dynamic evolution and local energy distribution characteristics. Avoid the uniform operation defects of traditional isotropic convolution, through the physical interpretable dimension decoupling, reduce the parameter redundancy while enhancing the feature discriminability, improve the unmanned aerial vehicle radio frequency fingerprint identification accuracy.
[0045] The present application solves the small sample generalization bottleneck of unmanned aerial vehicle radio frequency fingerprint identification through the collaborative design of spatiotemporal feature decoupling architecture and meta-learning optimization mechanism. The present application has the following beneficial effects:
[0046] First, the core component TFAnisotropic Block of the lightweight network LightWeightTFNet is proposed, and the anisotropic convolution kernel (5x1 time domain kernel and 1x7 frequency domain kernel) is used to separate and extract the transient fluctuation features and stable spectrum mode of the radio frequency signal, which significantly improves the physical interpretability of the time-frequency features and reduces the parameter redundancy; Second, the dynamic channel attention mechanism is introduced, which highlights the discriminative time-frequency area and suppresses noise interference through adaptive weighting; Third, the multi-scale pyramid structure is constructed, and efficient cross-scale feature interaction is realized through cross-layer residual connection, hierarchical down-sampling and channel expansion; Fourth, the model-independent meta-learning (MAML) framework is adopted, and the initialization parameters sensitive to task distribution are obtained through task-driven double-layer optimization (inner ring task adaptation and outer ring meta optimization), realizing the few-shot rapid adaptation ability only requiring a small amount of gradient update.
[0047] In addition, in the preprocessing stage, the original IQ signal is converted into a time-frequency diagram through short-time Fourier transform, providing a compatible two-dimensional convolution input carrier for feature extraction.
[0048] In summary, through the collaborative mechanism of time-frequency feature decoupling architecture and meta-learning optimization, the generalization and robustness of unmanned aerial vehicle radio frequency fingerprint identification are significantly improved.
[0049] Aiming at the feature confusion problem caused by the coupling of time domain transient fluctuation and frequency domain steady mode of radio frequency signals, a heterogeneous combination of 5x1 time domain convolution kernel and 1x7 frequency domain convolution kernel is adopted to capture short-time dynamic evolution and local energy distribution features respectively. This design avoids the uniform operation defects of traditional isotropic convolution, enhances the feature discriminability while reducing parameter redundancy through physically interpretable dimension decoupling, and directly deals with the noise interference problem in complex electromagnetic environment.
[0050] The channel attention mechanism is introduced, the global average pooling and the bottleneck layer are adaptively learned time-frequency region weight, so that the model focuses on the key discriminative region, the noise interference frequency band, and the robustness of the low signal-to-noise ratio scene is improved.
[0051] Based on the double-layer optimization framework of MAML, the task-specific adaptation is realized through single-step gradient update (inner loop) on the support set, and the meta-parameter update is driven by the query set loss (outer loop). This mechanism enables the model to obtain initialization parameters sensitive to task distribution, and only 5 samples are required to achieve an identification accuracy of 82.22%, thereby fundamentally solving the generalization problem caused by data scarcity.
[0052] The multi-scale pyramid structure integrates shallow details and deep semantics through cross-layer residual connection, so that the framework can achieve accuracy beyond complex models with only 4.3% of the parameter quantity of ResNet18, meeting the deployment requirements of resource-constrained scenarios.
[0053] When the unmanned aerial vehicle radio frequency fingerprint identification method provided by the application is applied, the unmanned aerial vehicle radio frequency fingerprint identification method provided by the application can be executed without Figure 1 The execution order of each step can be determined as required, and the application does not limit this.
[0054] The above is the unmanned aerial vehicle radio frequency fingerprint identification method provided by one or more embodiments of the application, based on the same idea, the application also provides a corresponding unmanned aerial vehicle radio frequency fingerprint identification device, as shown in Figure 5 .
[0055] Figure 5 The unmanned aerial vehicle radio frequency fingerprint identification device provided by the application is a schematic diagram, which comprises: The acquisition module 201 is used for acquiring the radio frequency signal of the unmanned aerial vehicle and determining the time-frequency diagram corresponding to the radio frequency signal; The convolution module 202 is used for inputting the time-frequency diagram into the radio frequency fingerprint identification network, performing time domain convolution through the sliding of the first convolution kernel of the radio frequency fingerprint identification network along the time axis of the time-frequency diagram, extracting the time domain features of the time-frequency diagram, performing frequency domain convolution through the sliding of the second convolution kernel of the radio frequency fingerprint identification network along the frequency axis of the time-frequency diagram, extracting the frequency domain features of the time-frequency diagram, and completing the anisotropic convolution of time-frequency; the size of the first convolution kernel is different from that of the second convolution kernel; The identification module 203 is configured to splice the time domain features and the frequency domain features to obtain spliced features, extract the radio frequency fingerprint features of the UAV through multi-scale time-frequency anisotropy convolution in the radio frequency fingerprint identification network according to the spliced features, and determine the UAV model according to the radio frequency fingerprint features.
[0056] The specific limitations of the UAV radio frequency fingerprint identification device can be referred to the limitations of the UAV radio frequency fingerprint identification method in the above, which will not be described here. The modules in the above UAV radio frequency fingerprint identification device can be realized by software, hardware and combinations thereof in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.
[0057] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the above Figure 1 The UAV radio frequency fingerprint identification method is provided.
[0058] The application further provides a computer device, which comprises a processor, an internal bus, a network interface, a memory and a non-volatile memory at the hardware level, and can further comprise other hardware required by other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs, so as to implement the above Figure 1 The UAV radio frequency fingerprint identification method is provided.
[0059] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments can be included. In the embodiments of the present application, any reference to the memory, storage, database or other medium can include at least one of the non-volatile and volatile memories. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory. The volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0060] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they should be considered to be within the scope of the present application.
Claims
1. A method for radio frequency fingerprint recognition of unmanned aerial vehicles, characterized in that, include: Acquire the radio frequency signal of the drone and determine the corresponding time-frequency diagram; The time-frequency graph is input into the radio frequency fingerprinting network. The first convolutional kernel of the radio frequency fingerprinting network slides along the time axis of the time-frequency graph to perform time-domain convolution and extract the time-domain features of the time-frequency graph. The second convolutional kernel of the radio frequency fingerprinting network slides along the frequency axis of the time-frequency graph to perform frequency-domain convolution and extract the frequency-domain features of the time-frequency graph, thus completing the time-frequency anisotropic convolution. The size of the first convolutional kernel is different from that of the second convolutional kernel. The time-domain features and frequency-domain features are concatenated to obtain the concatenated features. Based on the concatenated features, the radio frequency fingerprint features of the UAV are extracted through multi-scale time-frequency anisotropic convolution in the radio frequency fingerprinting network, so as to determine the UAV model based on the radio frequency fingerprint features.
2. The UAV radio frequency fingerprinting method as described in claim 1, characterized in that, The first convolution kernel is a 5×1 convolution kernel, and the second convolution kernel is a 1×7 convolution kernel.
3. The UAV radio frequency fingerprinting method as described in claim 1, characterized in that, The step of extracting the radio frequency fingerprint features of the UAV based on the splicing features through multi-scale time-frequency anisotropic convolution specifically includes: The spliced features are first reduced in dimensionality and then increased in dimensionality to determine the weight of each channel; the spliced features are then weighted by channel weights to obtain a weighted feature map. The radio frequency fingerprint features of the UAV are extracted by multi-scale time-frequency anisotropic convolution based on the weighted feature map.
4. The UAV radio frequency fingerprinting method as described in claim 3, characterized in that, The step of extracting the radio frequency fingerprint features of the UAV based on the weighted feature map through multi-scale time-frequency anisotropic convolution specifically includes: The weighted feature map is input into a multi-scale pyramid structure, and the radio frequency fingerprint features of the UAV are extracted through multi-scale time-frequency anisotropic convolution. The multi-scale pyramid structure includes a three-layer feature extraction structure. The first layer includes a regular convolutional layer and a time-frequency anisotropic convolutional layer. The second and third layers both include a downsampled time-frequency anisotropic convolutional layer and a time-frequency anisotropic convolutional layer. Cross-layer residual connections are provided between the various convolutional layers.
5. The UAV radio frequency fingerprinting method as described in claim 3, characterized in that, Training the radio frequency fingerprint recognition network specifically includes: The radio frequency signals of the drones are acquired as training samples. Each drone model is identified as a recognition task. A preset number of target recognition tasks are randomly sampled from the recognition tasks corresponding to the training samples. The training samples corresponding to each target recognition task are divided into a support set and a query set. For each target recognition task, the support set loss of the radio frequency fingerprinting network is determined by the support set of the target recognition task, and the radio frequency fingerprinting network is updated step by step to obtain the task parameters corresponding to the target recognition task. Based on the task parameters corresponding to the target recognition task, the query set loss of the radio frequency fingerprint recognition network is determined through the query set of the target recognition task. The radio frequency fingerprinting network is trained with the goal of minimizing the query set loss of the target identification task.
6. The UAV radio frequency fingerprinting method as described in claim 1, characterized in that, The determination of the time-frequency diagram corresponding to the radio frequency signal specifically includes: The radio frequency signal is segmented into frames using the Hanning window, and short-time Fourier transform is performed frame by frame to obtain the spectrum of the radio frequency signal. The magnitude of the radio frequency signal's spectrum is determined and spliced along the time axis to obtain the corresponding time-frequency diagram of the radio frequency signal.
7. A radio frequency fingerprint recognition device for unmanned aerial vehicles, characterized in that, include: The acquisition module is used to acquire the radio frequency signals of the drone and determine the corresponding time-frequency diagram of the radio frequency signals; The convolution module is used to input the time-frequency map into the radio frequency fingerprinting network. It performs time-domain convolution by sliding the first convolution kernel of the radio frequency fingerprinting network along the time axis of the time-frequency map to extract its time-domain features, and performs frequency-domain convolution by sliding the second convolution kernel of the radio frequency fingerprinting network along the frequency axis of the time-frequency map to extract its frequency-domain features, thus completing time-frequency anisotropic convolution. The size of the first convolution kernel is different from that of the second convolution kernel. The identification module is used to concatenate time-domain features and frequency-domain features to obtain concatenated features. Based on the concatenated features, the radio frequency fingerprint features of the UAV are extracted through multi-scale time-frequency anisotropic convolution in the radio frequency fingerprinting network, so as to determine the UAV model based on the radio frequency fingerprint features.
8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method as described in any one of claims 1 to 6.
9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 6.
Citation Information
Cited By
Unmanned aerial vehicle radio frequency fingerprint identification method and device, medium and equipment
CN121545188A
ASM-oriented radio frequency fingerprint identification method and system
CN121580172A
