A UAV RFID fingerprinting method based on feature fusion of dual-stream convolutional networks

Abstract

This application provides a method for UAV radio frequency fingerprinting based on feature fusion using a dual-stream convolutional network. The method includes: first, acquiring the short-time Fourier transform (STFT) time-frequency feature map of the radio frequency signal emitted by the UAV under test, and determining the corresponding time-frequency grayscale map and time gradient map; then, sequentially using the following components of the UAV radio frequency fingerprinting network: a static flow feature extraction network to extract features from the time-frequency grayscale map, obtaining static spatial texture features; a dynamic flow feature extraction network to extract features from the time gradient map, obtaining time-series dynamic change features; a fusion layer element-wise adding and fusing the static spatial texture features and the time-series dynamic change features to obtain fused features; and a classification layer predicting the category of the fused features to determine the category of the UAV under test. This application improves the accuracy of UAV radio frequency fingerprinting by achieving complementarity of dual-stream features and effectively controlling the number of model parameters.

Description

Technical Field

[0001] This application relates to the field of radio frequency identification for unmanned aerial vehicles (UAVs), and includes, but is not limited to, a UAV radio frequency fingerprinting method based on feature fusion of a two-stream convolutional network. Background Technology

[0002] In recent years, with the rapid development of microelectronics and communication technologies, unmanned aerial vehicle (UAV) systems have experienced explosive growth and application in both civilian and military fields. In civilian applications, UAVs are widely used in aerial photography, logistics delivery, agricultural plant protection, and power line inspection, greatly improving social productivity. In military applications, UAVs, as an important component of modern information warfare, have evolved from simple reconnaissance platforms into tactical nodes with multiple functions, including reconnaissance and strike capabilities, electronic warfare, and communication relay. However, the low, slow, and small characteristics of UAVs make them highly susceptible to breaching the detection blind spots of traditional air defense radars, leading to increasingly serious threats such as unauthorized flights, privacy violations, and public safety threats. Especially in complex electromagnetic environments and battlefield confrontations, how to quickly and accurately identify non-cooperative UAVs has become a critical issue that urgently needs to be addressed in low-altitude airspace control and counter-UAV warfare. Summary of the Invention

[0003] To address the above issues, this application provides a UAV RF fingerprinting method based on dual-stream convolutional network feature fusion, aiming to improve the accuracy of UAV RF fingerprinting while achieving complementarity of dual-stream features and effectively controlling the number of model parameters.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a method for UAV radio frequency fingerprinting based on feature fusion of a dual-stream convolutional network. The method includes: acquiring the short-time Fourier transform (STFT) time-frequency feature map of the radio frequency signal emitted by the UAV under test, and determining the corresponding time-frequency grayscale map and time gradient map; using the static flow feature extraction network in the UAV radio frequency fingerprinting network to extract features from the time-frequency grayscale map to obtain static spatial texture features; using the dynamic flow feature extraction network in the UAV radio frequency fingerprinting network to extract features from the time gradient map to obtain time-series dynamic change features; using the fusion layer in the UAV radio frequency fingerprinting network to perform element-wise addition and fusion of the static spatial texture features and the time-series dynamic change features to obtain fused features; and using the classification layer in the UAV radio frequency fingerprinting network to predict the category of the fused features to obtain the category of the UAV under test.

[0005] In some embodiments, both the static flow feature extraction network and the dynamic flow feature extraction network consist of four cascaded basic blocks with feature resolution halved at each stage. Each basic block contains, in sequence, a convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer.

[0006] In some embodiments, acquiring the short-time Fourier transform (STFT) time-frequency feature map of the radio frequency signal emitted by the UAV under test, and determining the time-frequency grayscale map and time gradient map corresponding to the STFT time-frequency feature map, includes: performing STFT conversion on the baseband received signal corresponding to the radio frequency signal to obtain the STFT time-frequency feature map; performing grayscale and normalization processing on the STFT time-frequency feature map to obtain the time-frequency grayscale map; and performing a first-order difference operation on the STFT time-frequency feature map in the time dimension to obtain the time gradient map.

[0007] In some embodiments, a static spatial texture feature is extracted from a time-frequency grayscale image using a static flow feature extraction network in a UAV radio frequency fingerprinting network. This includes: processing the input time-frequency grayscale image using a convolutional layer in the first-level base block of the static flow feature extraction network to obtain time-frequency texture features; performing a nonlinear transformation on the time-frequency texture features using a batch normalization layer and a ReLU activation function to obtain time-frequency texture activation features; downsampling the time-frequency texture activation feature image using a max pooling layer to obtain first-level time-frequency texture pooling features; and repeating the above steps using the first-level time-frequency texture pooling features as input and the remaining cascaded base blocks in the static flow feature extraction network to obtain static spatial texture features.

[0008] In some embodiments, a dynamic flow feature extraction network in an unmanned aerial vehicle (UAV) radio frequency fingerprinting network is used to extract features from a temporal gradient map to obtain temporally dynamic change features. This includes: processing the input temporal gradient map using a convolutional layer in the first-level base block of the dynamic flow feature extraction network to obtain temporal gradient features; performing a nonlinear transformation on the temporal gradient features using a batch normalization layer and a ReLU activation function to obtain temporal gradient activation features; downsampling the temporal gradient activation feature map using a max pooling layer to obtain first-level temporal gradient pooling features; and repeating the above steps using the first-level temporal gradient pooling features as input and the remaining cascaded base blocks in the dynamic flow feature extraction network to obtain temporally dynamic change features.

[0009] In some embodiments, the construction process of the UAV radio frequency fingerprinting network includes: acquiring the DroneRF dataset and preprocessing the DroneRF dataset to obtain a preprocessed dataset; using the preprocessed dataset, iteratively training and evaluating the performance of a lightweight two-stream convolutional network using label smoothing cross-entropy loss to obtain the UAV radio frequency fingerprinting network; wherein, the lightweight two-stream convolutional network is a network built on an architecture based on parallel feature extraction of static and dynamic streams combined with element-wise additive feature fusion.

[0010] The beneficial effects of the technical solutions provided in this application include at least the following: The UAV RF fingerprinting method based on dual-stream convolutional network feature fusion provided in this application utilizes a dual-stream parallel architecture of static STFT time-frequency feature maps and dynamic temporal gradient maps. The static stream in the UAV RF fingerprinting network captures the energy distribution and spectral contour of the STFT time-frequency feature maps, while the dynamic stream suppresses static background noise and enhances signal transient changes through temporal gradient map operations, thereby improving the signal-to-noise ratio and discriminability at the feature level. Simultaneously, a lightweight feature fusion strategy based on element-level addition is proposed in the UAV RF fingerprinting network. Thus, through dual-stream feature complementarity and temporal gradient enhancement, key problems in UAV RF fingerprinting can be effectively solved. Furthermore, the lightweight optimization strategy significantly reduces the number of model parameters and computational complexity while maintaining recognition accuracy, providing a complete theoretical and engineering solution for real-time, edge-end recognition of UAV RF signals. In other words, by achieving dual-stream feature complementarity and effectively controlling the number of model parameters, the accuracy of UAV RF fingerprinting can be improved.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the technical solutions provided in the embodiments of this application. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a UAV radio frequency fingerprinting method based on feature fusion of a two-stream convolutional network, provided for an embodiment of this application; Figure 2 A schematic diagram of the time-domain waveform, STFT time-frequency characteristic diagram, and time-frequency grayscale diagram of the radio frequency signal provided in the embodiments of this application; Figure 3A schematic diagram illustrating the composition of basic blocks in a lightweight convolutional backbone network provided in this application embodiment; Figure 4 A schematic diagram illustrating the framework for drone identification using the drone radio frequency fingerprinting network provided in this application; Figure 5 This is a schematic diagram illustrating the evolution of the feature dimensions of LDSC-Net provided in an embodiment of this application. Figure 6 A schematic diagram showing the overall recognition accuracy results obtained from the ablation experiments provided in this application; Figure 7 A schematic diagram showing the classification accuracy results obtained from the ablation experiment provided in this application; Figure 8 A visual illustration of the complementary features of the recognition effect of dynamic and static flow pairs provided in this application; Figure 9 A diagram illustrating the performance comparison of different feature fusion strategies provided in this application; Figure 10 A schematic diagram of the confusion matrix for classification recognition under different feature fusion strategies provided in this application; Figure 11A A comparative schematic diagram of the original STFT spectra corresponding to different UAV radio frequency signal samples provided in this application; Figure 11B A comparative schematic diagram of the static flow corresponding feature maps extracted for different UAV radio frequency signal samples provided in this application; Figure 11C A comparative schematic diagram of the radio frequency signal gradient maps corresponding to different UAV radio frequency signal samples provided in this application; Figure 11D A comparative schematic diagram of the dynamic flow corresponding feature maps extracted for different UAV radio frequency signal samples provided in this application; Figure 12 This application provides a visualization diagram of the activation of dynamic and static dual-stream features. Figure 13A A partial contrast diagram of the original time-frequency plots corresponding to the different UAV radio frequency signal samples provided in this application; Figure 13B A schematic diagram of local contrast of the differential gradient maps corresponding to different UAV radio frequency signal samples provided in this application; Figure 13C A comparative schematic diagram of pixel distribution histograms corresponding to different UAV radio frequency signal samples provided in this application; Figure 14 A diagram illustrating the comparison of recognition accuracy between LDSC-Net and other frameworks on the DroneRF dataset provided in this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0015] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0016] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0017] To address the problem of drone detection and identification, current mainstream technologies include radar detection, photoelectric identification, acoustic identification, and radio frequency (RF) identification. Among these, radar detection produces weak echoes from low-altitude, slow-moving, and small targets, which are easily obscured by clutter; photoelectric identification is significantly affected by weather, lighting conditions, and obstructions; and acoustic identification has a limited effective range. In contrast, RF identification technology intercepts the communication link signal between the drone and its remote controller, utilizing the "radio frequency fingerprint" formed by subtle differences in the drone's hardware to achieve individual identification. It boasts significant advantages such as long operating range, good concealment, and no line-of-sight limitations, making it a current hot research area in drone identification.

[0018] Early RF fingerprinting methods relied primarily on manually designed features combined with traditional machine learning classifiers. However, manually designed features have limited expressive power and poor robustness in complex electromagnetic environments such as low signal-to-noise ratios and multipath interference. With the rise of deep learning, Convolutional Neural Networks (CNNs) have been widely applied to RF signal recognition tasks due to their powerful automatic feature learning capabilities. Meanwhile, related literature has publicly released a large-scale drone RF signal dataset (DroneRFa) for detecting low-altitude drones. This dataset contains 454 RF recordings from various drone models under different flight modes and implements a 2, 4, and 10-class three-level classification task based on deep neural networks, providing an important public benchmark for subsequent research. Furthermore, addressing the needs of RF signal feature extraction and aircraft type identification, related literature has proposed a CNN|Triplet-DNN heterogeneous network model, as well as drone detection and identification methods based on multidimensional signal features and Mel-frequency cepstral coefficients. The latter employs a lightweight gated recurrent unit (GRU) architecture, effectively reducing model complexity while maintaining recognition accuracy.

[0019] Despite significant progress made by machine learning methods in the field of drone radio frequency identification, existing research still faces two prominent problems: 1. The problem of feature singularity: Most methods only use the static spectral features of the Short-Time Fourier Transform (STFT) time-frequency image as input, focusing on learning the energy distribution and spectral contour of the signal, neglecting the subtle dynamic changes of RF fingerprints in the time dimension, resulting in insufficient feature discriminative power under low signal-to-noise ratio conditions. To address this problem, feature fusion methods have been introduced into the field of signal recognition. Relevant literature utilizes sparse autoencoders to perform feature-level fusion recognition of high-resolution range images, time-frequency images, and range-instantaneous Doppler images. The dual-stream convolutional network architecture proposed in related literature extracts complementary features through parallel spatial and temporal streams, achieving breakthrough progress in video action recognition and providing inspiration for RF recognition tasks.

[0020] 2. Model complexity issue: In pursuit of recognition accuracy, researchers have adopted deep backbone networks with a large number of parameters, such as ResNet and EfficientNet. The number of model parameters is usually in the millions or even tens of millions. The storage overhead and inference latency are difficult to meet the needs of tactical edge devices.

[0021] To address the above issues, this application provides a UAV RF fingerprinting method based on dual-stream convolutional network feature fusion. It designs a dual-stream parallel architecture of "static STFT time-frequency feature map + dynamic time gradient map." The static stream captures the energy distribution and spectral contour of the STFT time spectrum, while the dynamic stream suppresses static background noise and enhances transient signal changes through time gradient operations, thereby improving the signal-to-noise ratio and discriminability at the feature level. Simultaneously, a lightweight feature fusion strategy based on element-level addition is proposed, effectively controlling the number of model parameters while ensuring the complementarity of the dual-stream features, thus improving the accuracy of UAV RF fingerprinting on a lightweight model architecture.

[0022] Example 1: See Figure 1 This is a flowchart illustrating a UAV radio frequency fingerprinting method based on dual-stream convolutional network feature fusion provided in this application embodiment. This UAV radio frequency fingerprinting method based on dual-stream convolutional network feature fusion can be executed by an electronic device, such as a computer or server.

[0023] like Figure 1 As shown, the UAV radio frequency fingerprinting method based on feature fusion of a two-stream convolutional network includes: Step 101: Obtain the short-time Fourier transform (STFT) time-frequency characteristic map of the radio frequency signal emitted by the UAV under test, and determine the time-frequency grayscale map and time gradient map corresponding to the STFT time-frequency characteristic map.

[0024] In some embodiments, the receiving end of the receiver is designed to identify the drone under test based on the received radio frequency signal emitted by the drone under test. Here, step 101 can be implemented by steps 1011 to 1013. Figure 1 (not shown in the image) Step 1011: Perform STFT conversion on the baseband received signal corresponding to the radio frequency signal to obtain the STFT time-frequency characteristic map.

[0025] In some embodiments, in a drone radio frequency fingerprinting scenario, the receiver will process the radio frequency signal emitted by the drone under test. Perform down-conversion processing to obtain the baseband received signal. : Formula (1); in, It is Additive White Gaussian Noise (AWGN); The time variable of the signal; The hardware damage characteristics of the UAV under test, namely the radio frequency fingerprint, are caused by inherent hardware defects of the UAV under test, such as local oscillator frequency offset, in-phase / quadrature imbalance (I / Q) and nonlinear characteristics of power amplifier. These characteristics are usually very weak, nonlinear and hidden in the transient or steady-state segments of the radio frequency signal.

[0026] Because the hardware damage characteristics of the UAV under test, i.e., the radio frequency fingerprint (feature), often exhibit non-stationary characteristics in both the time and frequency domains, it is difficult to capture the joint time-frequency dependence by directly processing one-dimensional signals. Therefore, this application uses STFT to... Transform to the time-frequency domain to obtain the STFT time-frequency characteristic map. : Formula (2); in, Indicates the window function; The time shift parameter of the window function; For frequency variables; It is a complex exponential kernel function.

[0027] Step 1012: Perform grayscale and normalization processing on the STFT time-frequency feature map to obtain a time-frequency grayscale map.

[0028] In some embodiments, the above-described It not only includes The modulation information, and more importantly, the embedding of two types of heterogeneous features: transient on / off features on the time axis and steady-state spectral patterns on the frequency axis. These fingerprint features, caused by hardware damage to the UAV under test, are... These features typically manifest as specific texture distributions or asymmetrical singular regions, which differ significantly from the target objects in natural images, thus requiring specially designed network architectures for extraction.

[0029] Furthermore, to ensure the robustness and computational efficiency of the UAV radio frequency fingerprinting network proposed in this application, this application... Perform grayscale and normalization processing to... Convert to a single-channel grayscale image and uniformly resize to 128×128. Then, normalize the pixel values ​​to the [-1, 1] interval to obtain the time-frequency grayscale image. This normalization process effectively accelerates the convergence speed of the stochastic gradient descent optimizer. Additionally, refer to... Figure 2 Among them, 201 is the time-domain waveform of the radio frequency signal emitted by the UAV under test, 202 is the STFT time-frequency characteristic diagram, and 203 is the time-frequency grayscale diagram.

[0030] Step 1013: Perform a first-order difference operation on the STFT time-frequency feature map in the time dimension to obtain the time gradient map.

[0031] In some embodiments, for Performing a first-order difference operation in the time dimension yields the dynamic input features, i.e., the time gradient map. And zero-padding is used to maintain the feature map size and STFT time-frequency characteristics. Figure 1 To: Formula (3); in, For the time dimension gradient operator; The time-frequency characteristic diagram of STFT is shown in the first... The value at each time step.

[0032] The above operations are equivalent to a high-pass filter along the time axis, which can suppress low-frequency noise interference, improve the signal-to-noise ratio of the feature region, and enable the dynamic flow feature extraction network in the subsequent UAV RF fingerprinting network to focus on learning the temporal variation pattern of signal energy. Here, the first-order difference spectrum... The pixel values ​​can reflect the amplitude change rate of the STFT time-frequency feature map in the time dimension, and the time-series dynamic change law of the UAV radio frequency signal at different frequencies. This operation has two significant advantages: First, it has strong dynamic acquisition capability. The first-order difference can effectively remove the DC component and background static components in the signal, highlighting the change trend of the signal on the time axis; Second, it has the advantage of enhancing the signal-to-noise ratio. Assuming that the background noise is additive white Gaussian noise, it has weak correlation between adjacent time frames, while the signal components have strong time-domain correlation.

[0033] It should be noted that, This is a first-order difference spectrum.

[0034] Step 102: Use the static flow feature extraction network in the UAV radio frequency fingerprinting network to extract features from the time-frequency grayscale image to obtain static spatial texture features.

[0035] Step 103: Use the dynamic flow feature extraction network in the UAV radio frequency fingerprinting network to extract features from the time gradient map and obtain the time-series dynamic change features.

[0036] In some embodiments, both the static flow feature extraction network and the dynamic flow feature extraction network consist of four cascaded basic blocks with feature resolution halved at each stage. Each basic block contains, in sequence, a convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer.

[0037] Here, both the static flow feature extraction network and the dynamic flow feature extraction network are implemented using a structurally symmetric and parameter-independent lightweight convolutional backbone network. Each flow (lightweight convolutional backbone network) consists of four cascaded basic blocks, where each basic block is composed of... Figure 3 The diagram shows a convolutional layer (Conv2d), a batch normalization layer (BN), a ReLU activation function, and a max pooling layer (MaxPool2d), which are used to achieve efficient feature abstraction and dimensionality reduction with a limited number of parameters.

[0038] In some embodiments, step 102 can be implemented by steps 1021 and 1022. Figure 1 (not shown in the image) Step 1021: In the first-level base block of the static flow feature extraction network, the input time-frequency grayscale image is processed by a convolutional layer to obtain time-frequency texture features; the time-frequency texture features are nonlinearly transformed by a batch normalization layer and a ReLU activation function to obtain time-frequency texture activation features; the time-frequency texture activation feature map is downsampled by a max pooling layer to obtain the first-level time-frequency texture pooling features.

[0039] Step 1022: Using the first-level time-frequency texture pooling feature as input, extract the remaining cascaded basic blocks in the network using static flow features, and repeat the above steps to obtain static spatial texture features.

[0040] Similarly, step 103 above can be achieved through steps 1031 and 1032. Figure 1 (not shown in the image) Step 1031: In the first-level base block of the dynamic flow feature extraction network, the input temporal gradient map is processed by a convolutional layer to obtain temporal gradient features; the temporal gradient features are nonlinearly transformed by a batch normalization layer and a ReLU activation function to obtain temporal gradient activation features; the temporal gradient activation feature map is downsampled by a max pooling layer to obtain the first-level temporal gradient pooling features.

[0041] Step 1032: Using the first-level temporal gradient pooling features as input, extract the remaining cascaded basic blocks in the network using dynamic flow features, and repeat the above steps to obtain the temporal dynamic change features.

[0042] Following the description above, the first One basic block ( The input feature map with values ​​of 1, 2, 3, 4 can be described as follows: Here, if This is a time-frequency grayscale image, i.e., the [number]th [item]. The basic blocks are the first-level basic blocks in the static flow feature extraction network. =1), which obtains time-frequency texture features by using the convolutional layers in the first-level base blocks of the static flow feature extraction network and then using two-dimensional convolutional blocks with 3×3 convolutional kernels. : Formula (4); in, and Indicates the first The learnable weights and biases of the convolutional layers within each basic block, i.e., the learnable weights and biases of the convolutional layers in the first-level basic block of the static flow feature extraction network. Represent the space of real numbers; , , These represent the number of channels, height, and width of the input time-frequency grayscale image, respectively.

[0043] Then, the time-frequency texture features output by the convolutional layer are processed using formula (5). Batch normalization and ReLU activation are performed to obtain time-frequency texture activation features. Accelerate training and introduce nonlinearity: Formula (5); in, , These represent the ReLU activation operation and the batch normalization operation, respectively.

[0044] Finally, a 2×2 kernel max pooling layer with a stride of 2 is used to activate the time-frequency texture features. The spatial size is halved to obtain the first-level time-frequency texture pooling feature. .

[0045] Formula (6); in, This represents a two-dimensional max pooling operation; This represents the number of channels for the first-level time-frequency texture pooling feature.

[0046] Thus, the dual-branch feature extraction module in the UAV RF fingerprinting network consists of a static flow feature extraction network (static flow spatial feature extraction branch) and a dynamic flow feature extraction network (dynamic flow temporal feature extraction branch). The two branches use lightweight convolutional networks with identical structures and non-shared weights to extract two types of core features (static spatial texture features and temporal dynamic change features) of the RF signal emitted by the UAV under test. This achieves decoupling and accurate extraction of multi-dimensional features, and the feature dimension evolution of the two branches remains consistent, laying the dimensional foundation for subsequent feature fusion. The static flow spatial feature extraction branch takes a time-frequency grayscale image as input and aims to extract static spatial texture features from the grayscale image, including the frequency distribution pattern, modulation texture, and time-frequency domain contour of the radio frequency signal. These features are the basis for distinguishing radio frequency signals emitted by different UAVs. The dynamic flow temporal feature extraction branch is an innovative design for the temporal variation characteristics of UAV radio frequency signals. It improves upon the static flow spatial feature extraction branch's neglect of signal temporal features. It utilizes the feature that the gradient change of the signal reflects the dynamic temporal law. Through a convolutional network with a structure completely consistent with the static flow spatial feature extraction branch, it analyzes the temporal gradient map obtained by calculating the gradient in the time dimension of the STFT time-frequency feature map to achieve the representation of the temporal dynamic features of the temporal gradient map.

[0047] In this application, both the static flow spatial feature extraction branch and the dynamic flow temporal feature extraction branch adopt a convolutional neural network composed of four cascaded basic blocks with progressively halved feature resolution. It has a powerful spatial feature extraction capability and achieves layer-by-layer mining from low-level texture features to high-level abstract features through the combination of multi-layer convolution, batch normalization, activation functions and pooling.

[0048] Step 104: Using the fusion layer in the UAV radio frequency fingerprinting network, the static spatial texture features and the temporal dynamic change features are added and fused element by element to obtain the fused features.

[0049] In some embodiments, to fully utilize the complementarity between the static spatial texture features output by the static flow feature extraction network and the temporal dynamic change features output by the dynamic flow feature extraction network, the fusion layer in the UAV RFID network maps the dual-flow features (static spatial texture features and temporal dynamic change features) to a unified feature space. Here, the fusion layer in the UAV RFID network is key to achieving complementary enhancement of dual-branch features. This application follows the principles of lightweight and parameterless operation, adopting an element-wise addition fusion strategy, requiring no additional learnable parameters. While ensuring the fusion effect, it minimizes the number of parameters and computational complexity of the UAV RFID network, meeting the needs of edge deployment. Let the static spatial texture feature be... The time-series dynamic change characteristics are The fusion feature is : + Formula (7).

[0050] Step 105: Using the classification layer in the UAV radio frequency fingerprinting network, perform category prediction on the fused features to obtain the probability distribution of the category to which the UAV to be tested belongs.

[0051] In some embodiments, fusion features First, the feature vector is compressed to 1×1 using global average pooling in the classification layer. : Formula (8); in, This is a global average pooling operation; , For fusion features The number of pixel rows (height) and the number of pixel columns (width); , These are the indices of the pixel row number and the pixel column number, respectively. =1, 2, ..., ; =1, 2, ..., ; For fusion features In the line, number The values ​​in the column; This represents the number of feature channels to be fused.

[0052] Here, the obtained feature vector The data is then fed into a classifier consisting of two fully connected layers in the classification layer: Formula (9); in, and These are the weight matrices of the first fully connected layer and the second fully connected layer in a two-layer fully connected network, respectively. and These are the bias vectors for the first and second fully connected layers, respectively. For batch normalization operations; To modify the activation function of the linear unit; For Dropout regularization operations; This is the 128-dimensional hidden feature vector output by the first fully connected layer; For the classifier output dimensional prediction vector; This represents the total number of categories to which the drone under test belongs. Here, we take... The index corresponding to the maximum value in the index can be used to obtain the identification category of the drone under test, i.e., its category.

[0053] Here, after the classifier, you can also... The probability distribution of the category to which the tested drone belongs is obtained by normalization using the Softmax function. : Formula (10); in, The drone under test belongs to the first The predicted probability of a class.

[0054] In some embodiments, such as Figure 4 As shown, the UAV RF fingerprinting network is a lightweight dual-stream convolutional-network (LDSC-Net). It focuses on the time-frequency features of the RF signals emitted by the UAV, employing a dual-branch parallel feature extraction architecture consisting of static and dynamic streams combined with element-wise feature fusion. This architecture achieves deep mining of the spatial texture and temporal dynamic features of the RF signals, and accurately determines the UAV category through a lightweight classifier. LDSC-Net takes the time-frequency grayscale image and temporal gradient image corresponding to the STFT time-frequency feature map as input, constructing two parallel feature extraction branches: a static stream and a dynamic stream. The static stream directly performs convolution operations on the time-frequency grayscale image to capture the static structure of the signal in the time-frequency domain; the dynamic stream then uses a convolutional network with the same structure to extract the dynamic change features of the signal from the temporal gradient image. The features output from the two feature extraction branches are fused through element-wise addition. The fused features include an adaptive average pooling and fully connected classifier, ultimately outputting the probability distribution of the UAV's category (e.g., ...). Figure 4 As shown: Drone 1, Drone 2, Drone 3, Drone 4.

[0055] In this application, the core advantages of the LDSC-Net are: 1) explicitly separating and fully utilizing the static and dynamic information of the signal; 2) enhancing the saliency of dynamic features through temporal gradient preprocessing; 3) adopting a consistent lightweight design for all convolutional blocks to control the number of model parameters; and 4) achieving feature complementarity through element-level additive fusion without increasing the number of parameters.

[0056] Furthermore, the UAV RF signal recognition framework corresponding to LDSC-Net is based on theories such as time-frequency analysis, convolutional neural networks, and dual-stream feature fusion. It has completed a fully modular design from signal acquisition to category determination, tailored to the characteristics of UAV RF signals and engineering deployment requirements. Each module's operation is supported by clear mathematical formulas, and feature evolution follows clear internal rules. The entire feature evolution process is free of dimensional conflicts and information redundancy; each step of dimensional change follows strict mathematical calculation logic, ensuring the effectiveness of feature extraction, the rationality of feature fusion, and the accuracy of classification and recognition. Here, the feature dimension evolution of LDSC-Net (taking the static flow feature extraction network as an example) is as follows: Figure 5 As shown.

[0057] In some embodiments, the construction process of the UAV radio frequency fingerprinting network provided in the above embodiments can be achieved through the following steps: The first step is to obtain the DroneRF dataset and preprocess it to obtain the preprocessed dataset.

[0058] The second step involves using a preprocessed dataset and employing label-smoothed cross-entropy loss to iteratively train and evaluate the performance of a lightweight two-stream convolutional network, resulting in a UAV RF fingerprint recognition network.

[0059] Among them, the lightweight dual-stream convolutional network is a network built on an architecture based on parallel feature extraction of static and dynamic streams combined with element-wise feature fusion.

[0060] In some embodiments, the open-source dataset, namely the DroneRF dataset, contains 454 radio frequency signal records, covering background signals and signals from three different types of drones (AR, Bebop, and Phantom). The signal sampling frequency is 40MHz, each segment lasts 0.25s, and contains 1,000,000 sampling points. The DroneRF dataset is divided into two frequency bands: a low-frequency band (2400~2440MHz) and a high-frequency band (2440~2480MHz), each containing 227 records, for a total of 454 records. In this application, STFT is used to process these data to form a unified time-spectrum dataset, with each image saved at a resolution of 450×300. This process generates 5448 spectrograms, covering four radio frequency signal categories: three categories of drone radio frequency signals and one category of background signals. Then, the dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0061] In this application, during training, a label-smoothed cross-entropy loss as shown in Equation (11) is used to improve the model's generalization ability. This loss is minimized by the Adam optimizer to update all network parameters: Formula (11); in, This represents the total loss value. This refers to the batch size, which is the number of samples in a single training iteration. This represents the total number of categories to which drones belong; , These are the sample index and the category index, respectively. =1, 2, ..., , =1,2,..., ; Indicates the first The true label of each sample; It is an exponential function, that is, when = When the condition is met, the value is 1; otherwise, it is 0. Indicates the smoothing coefficient; , The first The smoothed target distribution and model prediction probability of each sample.

[0062] The UAV RF fingerprinting method based on feature fusion of a dual-stream convolutional network provided in this application first acquires the short-time Fourier transform (STFT) time-frequency feature map of the RF signal emitted by the UAV under test, and determines the corresponding time-frequency grayscale map and time gradient map. Then, the static flow feature extraction network in the UAV RF fingerprinting network is used to extract features from the time-frequency grayscale map to obtain static spatial texture features. The dynamic flow feature extraction network in the UAV RF fingerprinting network is used to extract features from the time gradient map to obtain time-series dynamic change features. The fusion layer in the UAV RF fingerprinting network is used to fuse the static spatial texture features and the time-series dynamic change features element-wise to obtain fused features. Finally, the classification layer in the UAV RF fingerprinting network is used to predict the category of the fused features to obtain the category of the UAV under test. This application utilizes a dual-stream parallel architecture combining a static STFT time-frequency feature map and a dynamic time gradient map. The static stream in the UAV RF fingerprinting network captures the energy distribution and spectral contour of the STFT time-frequency feature map, while the dynamic stream suppresses static background noise and enhances signal transients through time gradient map computation, thereby improving signal-to-noise ratio and discriminability at the feature level. Simultaneously, a lightweight feature fusion strategy based on element-level addition is proposed for the UAV RF fingerprinting network. Thus, through dual-stream feature complementarity and time gradient enhancement, key issues in UAV RF identification can be effectively addressed. Furthermore, the lightweight optimization strategy significantly reduces the number of model parameters and computational complexity while maintaining recognition accuracy, providing a complete theoretical and engineering solution for real-time, edge-end recognition of UAV RF signals. In other words, by achieving dual-stream feature complementarity and effectively controlling the number of model parameters, the accuracy of UAV RF fingerprinting can be improved.

[0063] The following description, based on specific simulation experiments and result analysis data, illustrates the above-mentioned UAV radio frequency fingerprinting method based on feature fusion of dual-stream convolutional networks. However, it is worth noting that the simulation experiments and result analysis data are only for better illustrating this application and do not constitute an undue limitation on this application.

[0064] To verify the lightweight characteristics and classification performance of the UAV RF fingerprinting method based on dual-stream convolutional network feature fusion provided in this application, namely LDSC-Net, in the UAV RF fingerprinting task, this application conducts systematic experiments based on publicly available UAV RF fingerprint datasets, evaluating it from dimensions such as overall performance verification, ablation experiments, comparative experiments, and robustness analysis. Here, the publicly available source dataset DroneRF is used to evaluate the proposed LDSC-Net. The experiments are implemented on a server based on a PyTorch-based Nvidia RTX 5060 GPU platform.

[0065] A. Ablation Experiments: To verify the necessity of the dual-stream architecture proposed in LDSC-Net and the complementarity of static and dynamic features, a systematic ablation experiment was designed and conducted. Under the same training / validation / test set partitioning, data preprocessing procedures, and hyperparameter settings, the performance of the following three models was compared: First, the Static Only model, which contains only the static feature extraction stream, with the input being the time-frequency grayscale image corresponding to the STFT time-frequency feature map; its network structure is completely consistent with the static stream branch in LDSC-Net. Second, the Dynamic Only model, which contains only the dynamic feature extraction stream, with the input being the time gradient map corresponding to the STFT time-frequency feature map; its network structure is completely consistent with the dynamic stream branch in LDSC-Net. Third, the complete dual-stream fusion model (Fused (Dual-Stream)), which includes both static and dynamic streams and uses an element-wise addition strategy for feature fusion.

[0066] The overall recognition accuracy obtained from the ablation experiment is as follows: Figure 6 As shown, the recognition accuracy of the three models for each type of drone is as follows: Figure 7 As shown, the feature complementarity of the three models in terms of recognition performance is as follows: Figure 8 As shown. Here, Figure 6 The ablation experiments shown demonstrate that Static Only achieved an accuracy of 90.22% on the test set, proving that the time-frequency grayscale image corresponding to the STFT time-frequency feature map contains effective UAV fingerprint features; Dynamic Only achieved an accuracy of 83.15%, indicating that the temporal gradient features of the STFT time-frequency feature map also have discriminative power. However, single-stream networks all have limitations. The LDSC-Net proposed in this application, i.e., Fused (Dual-Stream), achieves an accuracy of 95.65%, which is 5 percentage points higher than the best single-stream network.

[0067] Further analysis Figure 7 The three models shown exhibit improved recognition accuracy across all categories, with Fused (Dual-Stream) demonstrating superior performance across all categories (AR, Bebop, Phantom, and background). Furthermore, as... Figure 8As shown, in the test set, Only Static (also a static flow independent model) correctly classified only 57 samples, Only Dynamic (also a dynamic flow independent model) correctly classified only 18 samples, while Fused (Dual-Stream) correctly classified 528 samples. This fully demonstrates that static texture features and dynamic change features are highly complementary in a physical sense, and the fusion strategy can effectively integrate multi-dimensional information and improve the overall generalization ability of the model.

[0068] B. Comparative Analysis of Feature Fusion Strategies: The design of the dual-stream network in LDSC-Net is highly dependent on the feature fusion strategy. Here, under the same network backbone, the performance differences of four mainstream fusion strategies are compared, and the results are as follows: Figure 9 As shown; where, Figure 9 Part (a) is the Accuracy Comparison, Part (b) is the F1-Score Comparison, Part (c) is the Precision Comparison, and Part (d) is the Recall Comparison.

[0069] according to Figure 9 The experimental results shown in sections (a) to (d) demonstrate that the element-wise addition strategy achieves performance comparable to or even better than complex attention mechanisms. While contact doubles the number of channels and retains all information, it is prone to overfitting on small datasets, resulting in a decrease in test set accuracy. Attention, although capable of adaptively adjusting weights, introduces additional parameters, making it difficult to converge sufficiently with limited data. In contrast, element-wise addition not only achieves zero parameter increase and maintains extremely low computational complexity, but also implicitly forces the features of both branches to learn consistently on the same dimension. This simple and effective strategy perfectly aligns with the "lightweight" requirement of this paper.

[0070] It should be noted that the fusion methods include: Contact, Attention, and Max. Contact concatenates the two feature paths along the channel dimension; this strategy retains all information but doubles the feature dimensionality and subsequent computational cost. Attention introduces learnable scalar weights, and after Softmax normalization, performs a weighted sum of the two feature paths. This strategy allows the network to adaptively learn the relative importance of the two feature paths. The Max strategy takes the maximum value at corresponding positions of the two feature paths. This strategy tends to retain the feature with the strongest response at each position and has non-linear selection characteristics.

[0071] To further analyze the performance of the two-stream fusion model in fine-grained classification, this application plotted the classification normalized confusion matrix under different feature fusion strategies, such as... Figure 10 As shown; where, Figure 10 Part (a) is the normalized confusion matrix for the element-wise addition fusion strategy; part (b) is the normalized confusion matrix for the contact fusion strategy; part (c) is the normalized confusion matrix for the attention fusion strategy; and part (d) is the normalized confusion matrix for the maximum element fusion strategy. Overall, LDSC-Net maintains an accuracy rate of over 90% for each category, demonstrating superior classification consistency compared to other feature fusion strategies.

[0072] C. Feature Visualization and Analysis: To intuitively explore the feature learning mechanism within the two-stream network, the feature map output from the last convolutional layer is extracted for visualization and statistical analysis, such as... Figures 11A to 11D As shown, the images sequentially represent the original STFT spectrum, static flow extracted feature map, RF signal gradient map, and dynamic flow extracted feature map corresponding to different UAV RF signal samples; among them, Figures 11A to 11D The drone categories corresponding to (1) to (4) shown are, in order: AR, Bebop, Phantom, and background. Figures 11A to 11D The paper presents a comparison of feature maps of radio frequency signal samples emitted by different drones in the dual-stream feature extraction path. It can be observed that the feature activation region of the static stream is usually highly consistent with the original STFT time-frequency feature map, mainly resolving the region with a large proportion of signal energy. This region presents a large-area blocky distribution, mainly responding to the steady-state spectral energy of the signal; while the feature activation of the dynamic stream is manifested as local highlighting, precisely focusing on the time-varying jump region of the signal.

[0073] Figure 12 A visualization of the activation of dynamic and static dual-flow features; among which, Figure 12 Part (a) shows the mapping between the activation intensity of dynamic and static dual-stream features and the categories (AR, Bebop, Phantom, background), further quantifying the differences in feature extraction between dynamic and static dual-stream. Taking the scatter plot of the activation pattern by class as an example, the horizontal axis is the average activation intensity of the static stream, and the vertical axis is the average activation intensity of the dynamic stream. Figure 12Part (b) shows the activation ratio distribution for different categories, where the horizontal axis represents the dynamic / static activation ratio and the vertical axis represents density. Figure 12 Part (c) shows the mean activation by class, where the horizontal axis represents each class (AR, Bebop, Phantom, background) and the vertical axis represents the mean activation value. This indicates that samples from different classes exhibit different distribution trends in the scatter plot, proving that the two branches do indeed capture different feature dimensions. Figure 12 Part (d) is Feature Discriminability Analysis, specifically feature discrimination analysis based on the Fisher discriminant ratio. The horizontal axis represents, in order: Intra-class Variance, Inter-class Variance, and Fisher Discriminant; the vertical axis represents the value. It shows that the ratio of inter-class variance to intra-class variance for dynamic flow features is significantly higher than that for static flow. This indicates that although the absolute accuracy of dynamic flow may not be superior, its extracted features have higher potential for class discrimination. It is this high discriminant feature that compensates for the shortcomings of static flow under interference in the same frequency band.

[0074] Here, since the core of dynamic flow branching lies in the first-order differential operation in the time dimension, in order to further explore the time-frequency characterization mechanism of time gradient features, i.e., the influence of the first-order differential operation on the separability of radio frequency features, this application starts with the energy distribution characteristics of the STFT time-frequency feature map for quantitative analysis. For example... Figure 11A As shown, from the perspective of time-frequency characterization, the pixel energy distribution of the original STFT time-frequency feature map often exhibits severe background noise accumulation, resulting in a weakening of the contrast of the effective signal features. Figure 11A and Figure 11C As shown in the partial visual effect comparison, the background area of ​​the original STFT time-frequency feature map has a flat energy base; while after time difference operation, the static background is effectively filtered out, and only the edge contours where the signal frequency jumps or fluctuates with time are retained.

[0075] To quantitatively assess this contrast variation, this application introduces a local contrast gain metric. Considering that UAV RF signal energy typically exhibits bandpass concentration in the time-frequency domain, this application defines the center frequency band region of the STFT time-frequency feature map as the active feature region and the two edge regions as the background reference region. The average pixel energy of each region is calculated to obtain the local contrast of the STFT time-frequency feature map. : Formula (12); in, and Let these represent the set of pixel coordinates of the active feature region and the set of pixel coordinates of the background reference region within the STFT time-frequency feature map, respectively. and This represents the total number of pixels.

[0076] The local contrast of the STFT time-frequency feature map and time gradient map were calculated based on formula (12), as shown in Figure 13; where, Figures 13A to 13C The images, in order, show local contrast diagrams of the original time-frequency maps, local contrast diagrams of the differential gradient maps, and comparison diagrams of the pixel distribution histograms corresponding to different UAV radio frequency signal samples; among them, Figures 13A to 13C The drone categories corresponding to (1) to (4) shown are, in order: AR, Bebop, Phantom, and background. See also Figures 13A to 13C The experimental results show that after time difference, the local contrast of the time gradient map is generally improved by more than 1 dB. Figure 13C The results show that the pixel value distribution of the STFT time-frequency feature map is more dispersed, with a wider dynamic range; while the pixel values ​​of the temporal gradient map are highly concentrated near zero, and the dynamic range is compressed. This is because the temporal difference operation essentially acts as a high-pass spatial filter. By suppressing the low-frequency background floor in the STFT time-frequency feature map, it relatively amplifies the energy differences of high-frequency dynamic edges. Although the overall dynamic range of pixel values ​​in the temporal gradient map is compressed, the relative energy ratio between its effective signal features and background noise is significantly improved. This enhanced contrast in time-frequency representation makes it easier for subsequent convolutional layers to capture discriminative texture edges, thereby improving the robustness of the model in feature extraction against complex backgrounds.

[0077] D. Comparative Analysis with Existing State-of-the-Art Models: To comprehensively evaluate the effectiveness and lightweight advantages of the proposed LDSC-Net, it was compared with existing models such as ResNet-18, EfficientNetV1, MobileNetV2, ShuffleNetV2, and SqueezeNet. All models were trained and tested under the same experimental settings to ensure the fairness of the comparison. The experimental results are shown in Table 1. Figure 14 As shown.

[0078] Table 1 Performance Comparison of Different Models here, Figure 14 (1) to (6) are, in order: the confusion matrix corresponding to ResNet-18, EfficientNetV1, MobileNetV2, ShuffleNetV2, SqueezeNet, and LDSC-Net for different drone categories (AR, Bebop, Phantom, background). The rows represent the true labels, i.e. the real categories, and the columns represent the predicted labels, i.e. the predicted categories. As can be seen from the data in Table 1, ResNet-18 has a recognition accuracy of up to 95.11% and an F1 value of 0.96, but the model has as many as 11.17M parameters, which is too large and difficult to meet the deployment requirements of edge devices. EfficientNetV1 has a recognition accuracy slightly higher than ResNet-18, reaching 95.65%, and the F1 value is also 0.96, but its parameter count is still 7.03M, which is still difficult to deploy for drone terminal devices with limited resources. MobileNetV2 has a moderate number of parameters (2.23M), but its recognition accuracy is only 93.48% and its F1 score is 0.93, showing a significant performance gap compared to LDSC-Net. ShuffleNetV2, while reducing its parameter count to 1.26M, suffers a sharp drop in recognition accuracy to 89.86% and an F1 score of only 0.90, indicating that its lightweight design excessively sacrifices feature representation capabilities, failing to meet the high-precision recognition requirements of drone RF fingerprints. SqueezeNet exhibits better lightweight performance, with only 0.73M parameters, achieving a recognition accuracy of 94.75% and an F1 score of 0.95, but its performance is still slightly lower than LDSC-Net.

[0079] In comparison, the proposed LDSC-Net achieves a recognition accuracy of 95.65% and an F1 score of 0.96, performing on par with large models such as ResNet-18 and EfficientNetV1, and significantly outperforming MobileNetV2 and ShuffleNetV2. Its parameter count is only 0.58M, a 30% reduction compared to SqueezeNet, over 50% smaller than ShuffleNetV2, nearly 75% smaller than MobileNetV2, and over 90% smaller than ResNet-18 and EfficientNetV1. This result fully demonstrates that LDSC-Net achieves an extremely lightweight design while maintaining high recognition accuracy, effectively balancing the core contradiction of "accuracy-lightweight," and is more suitable for the real-time RF fingerprint recognition needs of drones on edge devices.

[0080] Based on the above description, this application proposes a lightweight dual-stream convolutional network feature fusion method to address the problems of low signal-to-noise ratio feature extraction and model lightweighting in UAV radio frequency signal identification, and verifies its effectiveness through systematic experiments. This framework achieves the collaborative extraction of static time-frequency texture features and dynamic temporal gradient features through an innovative design of a static-stream + dynamic-stream dual-branch approach. The temporal gradient operation introduced by the dynamic-stream branch is essentially a high-pass spatial filter, which relatively amplifies the energy differences at the dynamic edges of the signal by suppressing the low-frequency background floor in the time-frequency map. Employing a single-channel STFT input and an element-level additive fusion strategy, the number of model parameters is controlled to 0.58M, significantly reducing computational complexity. Furthermore, compared with four feature fusion strategies, element-level additive fusion achieves the best balance between parameter quantity and performance, avoiding the overfitting problem of splicing fusion and the parameter redundancy of the attention mechanism, thus verifying the rationality of the lightweight design. In summary, the lightweight dual-stream convolutional network proposed in this application effectively solves the key problems in UAV radio frequency identification through dual-stream feature complementarity and temporal gradient enhancement. At the same time, through a lightweight optimization strategy, it significantly reduces the number of model parameters and computational complexity while ensuring recognition accuracy, providing a complete theoretical and engineering solution for real-time, edge-end recognition of UAV radio frequency signals.

[0081] In other words, addressing the shortcomings of traditional RFID fingerprinting methods in feature extraction robustness under low signal-to-noise ratio (SNR) environments and the large parameter count of existing deep learning models which makes them unsuitable for tactical edge deployments, this application proposes a lightweight dual-stream convolutional network feature fusion method for UAV RFID. This method constructs a parallel processing architecture for static and dynamic streams, collaboratively extracting static time-frequency texture features and dynamic temporal gradient features of the signal, and employs an element-level addition fusion strategy to achieve feature complementarity. Specifically, the static stream uses the short-time Fourier transform time-frequency map as input and extracts spectral contour features through stacked convolutional modules; the dynamic stream introduces temporal gradient operations, suppressing static background noise through first-order difference and enhancing the time-frequency contrast of the signal's dynamic edges. Simulation results on the open-source dataset DroneRF verify the effectiveness of the proposed method. The dual-stream fusion network achieves a recognition accuracy of 95.65% on the test set, a 5 percentage point improvement over the single-stream network; the temporal gradient features improve the average SNR by more than 1 dB, significantly enhancing feature separability under low SNR environments; and the model parameter count is only 0.58M, meeting the real-time requirements of tactical edge devices. Furthermore, ablation experiments, feature visualization, and comparative analysis of fusion strategies show that the complementarity of static and dynamic features and lightweight design are the core factors for improving recognition performance.

[0082] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0083] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways.

[0085] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0086] The features disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0087] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A UAV radio frequency fingerprinting method based on dual-stream convolutional network feature fusion, characterized in that, The method includes: Obtain the short-time Fourier transform (STFT) time-frequency characteristic map of the radio frequency signal emitted by the UAV under test, and determine the time-frequency grayscale map and time gradient map corresponding to the STFT time-frequency characteristic map; By utilizing the static flow feature extraction network in the UAV radio frequency fingerprinting network, feature extraction is performed on the time-frequency grayscale image to obtain static spatial texture features; By utilizing the dynamic flow feature extraction network in the UAV radio frequency fingerprinting network, feature extraction is performed on the temporal gradient map to obtain the time-series dynamic change features; By utilizing the fusion layer in the UAV radio frequency fingerprinting network, static spatial texture features and temporal dynamic change features are fused element-wise to obtain fused features; By utilizing the classification layer in the UAV radio frequency fingerprinting network, the category of the fused features is predicted to obtain the category of the UAV to be tested.

2. The method according to claim 1, characterized in that, Both the static flow feature extraction network and the dynamic flow feature extraction network consist of four cascaded basic blocks with feature resolution halved at each stage. Each basic block contains, in sequence, a convolutional layer, a batch normalization layer, a ReLU activation function, and a max pooling layer.

3. The method according to claim 1, characterized in that, Obtain the short-time Fourier transform (STFT) time-frequency characteristic map of the radio frequency signal emitted by the UAV under test, and determine the corresponding time-frequency grayscale map and time gradient map of the STFT time-frequency characteristic map, including: STFT conversion is performed on the baseband received signal corresponding to the radio frequency signal to obtain the STFT time-frequency characteristic map; The STFT time-frequency feature map is subjected to grayscale and normalization processing to obtain a time-frequency grayscale map; A first-order difference operation is performed on the STFT time-frequency characteristic map in the time dimension to obtain the time gradient map.

4. The method according to claim 2, characterized in that, By utilizing the static flow feature extraction network in the UAV radio frequency fingerprinting network, feature extraction is performed on the time-frequency grayscale image to obtain static spatial texture features, including: In the first-level base block of the static flow feature extraction network, the input time-frequency grayscale image is processed by a convolutional layer to obtain time-frequency texture features; the time-frequency texture features are nonlinearly transformed by a batch normalization layer and a ReLU activation function to obtain time-frequency texture activation features; and the time-frequency texture activation feature map is downsampled by a max pooling layer to obtain the first-level time-frequency texture pooling features. Using the first-level time-frequency texture pooling features as input, the static flow features are used to extract the remaining cascaded basic blocks in the network. The above steps are repeated to obtain static spatial texture features.

5. The method according to claim 2, characterized in that, By utilizing the dynamic flow feature extraction network in the UAV radio frequency fingerprinting network, feature extraction is performed on the temporal gradient map to obtain temporal dynamic change features, including: In the first-level base block of the dynamic flow feature extraction network, the input temporal gradient map is processed by a convolutional layer to obtain temporal gradient features; the temporal gradient features are nonlinearly transformed by a batch normalization layer and a ReLU activation function to obtain temporal gradient activation features; and the temporal gradient activation feature map is downsampled by a max pooling layer to obtain the first-level temporal gradient pooling features. Using the first-level temporal gradient pooling features as input, the remaining cascaded basic blocks in the network are extracted using dynamic flow features. The above steps are repeated to obtain the temporal dynamic change features.

6. The method according to claim 1, characterized in that, The process of building a drone radio frequency fingerprinting network includes: Obtain the DroneRF dataset and preprocess it to obtain the preprocessed dataset; Using a preprocessed dataset and employing label-smoothed cross-entropy loss, a lightweight two-stream convolutional network is iteratively trained and its performance evaluated to obtain an unmanned aerial vehicle (UAV) radio frequency fingerprinting network. The lightweight two-stream convolutional network is a network built on an architecture based on parallel feature extraction of static and dynamic streams combined with element-wise feature fusion.

Images