Multi-unmanned aerial vehicle radio frequency signal-oriented type identification method, device and equipment, and storage medium

Abstract

The invention discloses a model identification method, device and equipment for multi-unmanned aerial vehicle radio frequency signals and a storage medium, and relates to the technical field of mode identification and classification, and the method comprises the steps: carrying out the preprocessing of the unmanned aerial vehicle radio frequency signals, and obtaining a radio frequency time-frequency graph; performing feature extraction on the radio frequency time-frequency graph through a multi-branch heterogeneous network model to obtain a first branch feature, a second branch feature and a third branch feature; performing global feature coding, branch attention generation and feature weighting on the features of the three branches through a cross-branch attention interaction module to obtain attention enhancement features; the attention enhancement features are refined through a local and global feature refining module, and fusion features are obtained; and inputting the fusion features into a full connection layer and an activation function for classification to obtain a multi-unmanned aerial vehicle model identification result. According to the invention, the accuracy of identifying the radio frequency signal type of the unmanned aerial vehicle and the capability of resisting complex electromagnetic environment interference can be improved in a multi-vehicle coexistence scene.

Description

Technical Field

[0001] This application relates to the field of pattern recognition and classification technology, and in particular to a method, apparatus, device and storage medium for aircraft type identification based on radio frequency signals of multiple unmanned aerial vehicles (UAVs). Background Technology

[0002] With the rapid opening of low-altitude airspace and the widespread use of drones, illegal drone activities such as unauthorized and excessive flights pose a serious threat to public safety, privacy protection, and critical infrastructure. To achieve effective regulation and countermeasures, a technological means is urgently needed to identify multiple drones in the air in real time with high precision.

[0003] Current approaches primarily employ methods based on a single deep learning model (such as a convolutional neural network (CNN) or a recurrent neural network (RNN)) to extract features and classify the radio frequency signals of a single UAV. These methods typically acquire clean signal samples under ideal channel conditions and assume that only a single UAV exists in the scene, using supervised learning to train a classifier to achieve UAV type identification.

[0004] However, in real-world, complex scenarios, multiple drones of different types often fly simultaneously, causing their radio frequency (RF) signals to overlap in both the time and frequency domains, resulting in mixed signals. Furthermore, the actual electromagnetic environment presents adverse factors such as multipath effects, co-channel interference, and noise disturbances, leading to a significant deterioration in signal quality. Therefore, improving the accuracy of drone RF signal identification and its resistance to interference from complex electromagnetic environments in multi-drone coexistence scenarios has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, device, and storage medium for aircraft type identification based on radio frequency signals of multiple drones, aiming to solve the technical problem of how to improve the accuracy of drone radio frequency signal type identification and the ability to resist interference from complex electromagnetic environments in multi-drone coexistence scenarios.

[0006] To achieve the above objectives, this application proposes a method for aircraft type identification based on radio frequency signals of multiple unmanned aerial vehicles (UAVs), the method comprising: The radio frequency signal of the UAV is preprocessed to obtain the radio frequency time-frequency diagram; Feature extraction is performed on the radio frequency time-frequency map using a multi-branch heterogeneous network model to obtain the first branch feature, the second branch feature, and the third branch feature. The attention-enhanced features are obtained by performing global feature encoding, branch attention generation, and feature weighting on the first branch features, the second branch features, and the third branch features through the cross-branch attention interaction module. The attention enhancement features are refined using local and global feature refinement modules to obtain fused features; The fused features are input into a fully connected layer of a DNN and a Softmax activation function for classification, resulting in multi-drone model identification results.

[0007] Furthermore, to achieve the above objectives, this application also proposes a device for identifying the type of multiple unmanned aerial vehicles (UAVs) based on their radio frequency signals, the device comprising: The preprocessing module is used to preprocess the UAV radio frequency signal to obtain the radio frequency time-frequency diagram; The feature extraction module is used to extract features from the radio frequency time-frequency map through a multi-branch heterogeneous network model to obtain first branch features, second branch features and third branch features; The feature enhancement module is used to perform global feature encoding, branch attention generation, and feature weighting on the first branch feature, the second branch feature, and the third branch feature through the cross-branch attention interaction module to obtain attention-enhanced features; The refinement module is used to refine the attention enhancement features through local and global feature refinement modules to obtain fused features; The aircraft model recognition module is used to input the fused features into the fully connected layer of the DNN and the Softmax activation function for classification, so as to obtain the recognition results of multiple UAV models.

[0008] Furthermore, to achieve the above objectives, this application also proposes a model identification device for multiple UAV radio frequency signals. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the model identification method for multiple UAV radio frequency signals as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for identifying the type of multiple UAV radio frequency signals.

[0010] One or more technical solutions proposed in this application have at least the following technical effects: First, the UAV radio frequency (RF) signal is preprocessed to obtain an RF time-frequency map, effectively suppressing out-of-band noise and retaining discriminative time-frequency fingerprints, providing high-quality input for subsequent identification. Second, a multi-branch heterogeneous network model is used to extract low-level, mid-level, and high-level features from the RF time-frequency map using the first, second, and third branches respectively. These three branches are complementary, capturing aircraft-specific information from different levels of abstraction. Then, a cross-branch attention interaction module is used to concatenate, reduce dimensionality, and perform nonlinear mapping on the features from the three branches, generating attention weights for each branch and weighted fusion to achieve dynamic complementarity and enhancement of important information between features. Next, a local and global feature refinement module refines the attention-enhanced features to obtain more discriminative and robust fused features. Finally, the fused features are fed into a fully connected DNN layer for high-level semantic mapping, and the aircraft category probability is output through a Softmax activation function to complete the final identification. This application, through multi-level feature collaborative modeling and refined feature optimization, can improve the accuracy of UAV RF signal aircraft identification and its resistance to complex electromagnetic interference in multi-aircraft coexistence scenarios. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an embodiment of the method for identifying the type of multiple UAVs based on radio frequency signals in this application. Figure 2 This is a schematic diagram of the Triplet-CNN structure provided in Embodiment 1 of the method for aircraft model identification based on radio frequency signals of multiple UAVs in this application; Figure 3 This is a schematic diagram of the CNN|Triplet-CRGM model provided in Embodiment 1 of the method for aircraft model identification based on radio frequency signals of multiple UAVs in this application; Figure 4 This is a flowchart illustrating Embodiment 2 of the method for identifying the type of multiple UAVs based on radio frequency signals in this application. Figure 5 This is a schematic diagram of the CRIFA module provided in Embodiment 2 of the method for identifying the type of multiple UAV radio frequency signals in this application; Figure 6This is a simplified flowchart illustrating the aircraft type identification method for multiple UAV radio frequency signals provided in Embodiment 2 of this application; Figure 7 This is a schematic diagram of the performance parameters of the CNN|Triplet-CRGM model ablation experiment based on the RF time-frequency map of a single UAV, which is the aircraft type identification method for multi-UAV radio frequency signals provided in Embodiment 2 of this application. Figure 8 This diagram illustrates the training performance parameters of a CNN|Triplet-CRGM model based on the RF time-frequency map of a single UAV, which is used for the aircraft type identification method based on the RF signals of multiple UAVs provided in Embodiment 2 of this application. Figure 9 This is a schematic diagram of the training performance parameters of the CNN|Triplet-CRGM model of the RF time-frequency map of a single UAV in a real flight scenario for the UAV model identification method for multi-UAV RF signals provided in Embodiment 2 of this application. Figure 10 This is a schematic diagram of the module structure of the aircraft model identification device for multiple UAV radio frequency signals according to an embodiment of this application; Figure 11 This is a schematic diagram of the hardware operating environment of the aircraft model identification method for multiple UAV radio frequency signals in the embodiments of this application.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0016] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0017] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or device identification system capable of realizing the above functions. The following description uses a device identification system as an example to illustrate this embodiment and the subsequent embodiments.

[0018] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0019] Based on this, embodiments of this application provide a method for aircraft type identification based on radio frequency signals of multiple UAVs, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the aircraft model identification method for multiple UAV radio frequency signals in this application.

[0020] In this embodiment, the aircraft type identification method for multiple UAV radio frequency signals includes steps S10 to S50: Step S10: Preprocess the UAV radio frequency signal to obtain the radio frequency time-frequency diagram; It should be noted that drone radio frequency (RF) signals refer to the electromagnetic wave signals actively emitted by drones through radio channels during flight or communication. Their operating frequencies are typically located in industrial, scientific, and medical bands or other licensed bands. They contain individual-specific information introduced by differences in hardware circuitry and modulation methods, serving as an "RF fingerprint" for identifying different brands or models of drones. An RF time-frequency diagram is a two-dimensional visual representation obtained by converting a one-dimensional RF signal through time-frequency analysis methods (e.g., short-time Fourier transform). The horizontal axis represents time, the vertical axis represents frequency, and pixel intensity (e.g., color depth) reflects the energy magnitude of the corresponding frequency component at that moment, intuitively displaying the dynamic distribution characteristics of the signal in both time and frequency dimensions.

[0021] As an example, the step of preprocessing the UAV radio frequency signal to obtain a radio frequency time-frequency map includes: filtering the UAV radio frequency signal to obtain a filtered signal; performing a short-time Fourier transform on the filtered signal according to a preset sampling rate and a preset window function to obtain a time-frequency matrix; extracting effective signal frequency bands from the time-frequency matrix, wherein the effective signal frequency bands are within a preset frequency range and the signal strength is greater than a preset strength threshold; truncating and performing time-frequency analysis on the effective signal frequency bands according to preset time segments to obtain a continuous initial time-frequency map; and standardizing the initial time-frequency map to obtain a radio frequency time-frequency map.

[0022] The preset sampling rate refers to the number of samples per second pre-set when digitally acquiring the filtered UAV RF signal. This ensures complete capture of frequency components within the target signal bandwidth, satisfying the Nyquist sampling theorem. The preset window function (e.g., Hanning window, Hamming window) is a weighting function applied to local time intervals of the signal before performing a short-time Fourier transform to reduce spectral leakage. Its function is to smooth signal truncation edges and improve the frequency domain resolution and stability of time-frequency analysis. The time-frequency matrix is ​​a two-dimensional complex or energy matrix obtained by dividing the filtered signal into frames and performing a short-time Fourier transform frame by frame. Each column corresponds to a time frame, and each row corresponds to a frequency point. Matrix elements represent the signal amplitude or energy at that time-frequency location. The effective signal band refers to the frequency range within the preset band in the time-frequency matrix where the signal strength exceeds a preset strength threshold. This represents the useful spectrum region that may contain real UAV communication or control signals, used to eliminate noise and irrelevant interference. The preset frequency band range refers to the target frequency interval pre-defined based on common UAV communication protocols and operating frequency bands. This limits the frequency search space for subsequent analysis, improves processing efficiency, and focuses on typical UAV activity frequency bands. The preset strength threshold is the energy criterion used to distinguish effective signals from background noise. Only when the energy of a frequency point in the time-frequency matrix exceeds this threshold is it considered a potentially effective UAV signal component. The preset time segment refers to the basic unit of dividing continuous radio frequency signals into fixed-length time windows. This is used to generate time-frequency map samples with a uniform time scale, facilitating subsequent model input and batch processing. The continuous initial time-frequency map refers to the original time-frequency image sequence, unnormalized or unformatted, formed by arranging the images in chronological order after performing time-frequency analysis on each preset time segment. This preserves the complete time-frequency structure of the signal within a local time period.

[0023] First, the drone identification system performs bandpass or lowpass filtering on the raw acquired UAV radio frequency signals to remove strong out-of-band interference and high-frequency noise, retaining useful frequency bands that may contain UAV communication characteristics. Then, the system resamples the filtered signal at a preset sampling rate and applies a preset window function to the signal, dividing it into frames and performing a short-time Fourier transform on each frame to generate a complete time-frequency matrix. Next, the system locates frequency regions within the preset frequency band and with energy above a preset intensity threshold in this time-frequency matrix, extracting the effective signal frequency bands to eliminate the influence of background noise and irrelevant signals. Subsequently, the system truncates the effective signal frequency bands into preset time segments of equal length and performs time-frequency analysis again on each time segment, forming a time-continuous and structurally consistent initial time-frequency map sequence. Finally, the system standardizes these initial time-frequency maps, including normalizing pixel values ​​to the [0, 1] interval and unifying image size, to adapt to the input requirements of subsequent deep neural networks, ultimately outputting a standardized radio frequency time-frequency map.

[0024] Step S20: Extract features from the radio frequency time-frequency map using a multi-branch heterogeneous network model to obtain first branch features, second branch features, and third branch features; It should be noted that the multi-branch heterogeneous network model refers to a deep neural network architecture composed of three structurally different and functionally complementary sub-networks operating in parallel. Each branch focuses on extracting features at different levels or dimensions of the RF time-frequency map. This heterogeneous design enhances the model's ability to express complex RF fingerprints and its robustness. The first branch features are the local spatial texture features extracted from the RF time-frequency map by the first branch. This branch uses multi-layer convolution and pooling operations to capture subtle structural patterns of the signal in the time-frequency plane, such as modulation sidebands, transient transitions, and other fingerprint information related to hardware implementation. The second branch features are the embedded features learned by the second branch, possessing intra-class compactness and inter-class separability. This branch optimizes the feature space distribution by comparing the relative distances between samples of the same and different models, improving the model's ability to distinguish similar models. The third branch features are the global semantic features generated by the third branch through high-order nonlinear mapping of the one-dimensional vector of the RF time-frequency map after global statistics or dimensionality reduction. This branch focuses on capturing the overall statistical characteristics and high-level discriminative information of the signal to supplement the shortcomings of the first two branches in global context modeling.

[0025] As an example, the multi-branch heterogeneous network model includes a first branch, a second branch, and a third branch; the step of extracting features from the radio frequency time-frequency map using the multi-branch heterogeneous network model to obtain the first branch features, the second branch features, and the third branch features includes: extracting low-level features from the radio frequency time-frequency map using a three-layer convolutional structure of the first branch to obtain the first branch features; extracting mid-level features from the radio frequency time-frequency map using a four-layer convolutional structure of the second branch to obtain the second branch features; and extracting high-level features from the radio frequency time-frequency map using a five-layer triplet convolutional network of the third branch to obtain the third branch features.

[0026] The first branch refers to the sub-network in the multi-branch heterogeneous network model specifically designed for extracting low-level local features from the radio frequency (RF) time-frequency map. This first branch utilizes three convolutional layers to extract low-level features from the time-frequency map and introduces a cross-entropy loss function to achieve gradient backpropagation. By measuring the difference between the predicted result and the true label, it minimizes the cross-entropy to encourage the model prediction to converge towards the true category. The expression for the cross-entropy loss function is: Where N is the number of samples in the batch, and C is the number of categories. y i,c For the sample i Corresponding category c The tag, p i,cPredict samples for the model i Category c The probability of.

[0027] The second branch refers to the sub-network in the model responsible for extracting the upper-level semantic features from the radio frequency time-frequency image. This second branch employs a four-layer convolutional design, combined with a center loss function, to optimize the distance between signal samples and their respective category centers, enhancing the clustering of similar features in the feature space. The category center formula is: in, Z i For category i The center vector, For category i The Middle j The feature vector after mapping each sample, |C i |For category i The number of samples is determined. Based on the class centers, the centralized location of each sample in the feature space is determined. To achieve both minimizing intra-class divergence and maximizing inter-class separation, a center vector correlation factor is introduced to increase the inter-class distance, based on the original center loss function definition. The improved center loss function is: Where N is the number of signal samples, x i For the first i The feature vector of each sample y i For the first i The categories of each sample, Z yi For the first y i The center vector of the class, where C is the total number of classes. Z i and Z j They are the first i Class and First j The center vector of the class, The weight hyperparameters for the penalty term are used to measure the center loss and the inter-class distance penalty term.

[0028] The third branch refers to the sub-network in the model used to extract high-level discriminative features from the Radio Frequency (RF) time-frequency map. This third branch employs Triplet-CNN to extract RF time-frequency map features. Its feedforward network consists of five cascaded convolutional and pooling layers, progressively mining fine-grained features from the RF time-frequency map. Pooling layer downsampling reduces computational complexity and suppresses overfitting. Optimizing the branch network hyperparameters using the triplet loss function improves the model's ability to extract features from various device models. The expression for Triplet Loss is: in, This represents the branching feature of positive samples in triplet pairs. Branch features for triplet anchor point samples This represents the negative sample branching feature of triplet. for L 2-norm square, The sample number. The total number of samples; The interval threshold is used to avoid extreme values ​​from interfering with the feature space distance and to ensure that the model continues to learn.

[0029] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the Triplet-CNN structure provided in Embodiment 1 of the method for model identification of multiple UAV radio frequency signals in this application. The Triplet-CNN is composed of multiple functional modules connected sequentially from top to bottom. At the top is the triplet sample pair input layer, which receives three input samples: an anchor sample, a positive sample (of the same UAV model as the anchor), and a negative sample (of a different UAV model than the anchor), as input data for model training. Below this is the convolutional neural network layer, which consists of multiple layers of convolution, pooling, and activation functions. It is responsible for extracting features from the input triplet samples and generating corresponding embedding representations. Next is the distance layer, which calculates the distance between the anchor sample and the positive sample, and the distance between the anchor sample and the negative sample, to quantify their similarity in the feature space. Then comes the discriminant layer, which determines whether the triplet constraint relationship is satisfied based on the distance value output by the distance layer by setting a loss function constraint condition, i.e., the distance of the positive sample should be less than the distance of the negative sample. Finally, there is the output layer, which outputs the final discrimination result or loss signal and feeds it back to the previous layer for parameter updates. In addition, the figure also includes a feedback path from the discriminant layer to the convolutional neural network layer, labeled as triple loss. This means that during training, the triple loss function will backpropagate based on the output of the discriminant layer to optimize the weight parameters of the convolutional neural network layer, thereby achieving the goal of bringing similar samples closer and dissimilar samples further apart in the feature space.

[0030] As an example, the multi-branch heterogeneous network model includes a first branch, a second branch, and a third branch. The first branch uses a cross-entropy loss function, the second branch uses a center loss function, and the third branch uses a triplet loss function. The training steps of the multi-branch heterogeneous network model include: preprocessing radio frequency signal samples from multiple types of UAVs and dividing them into training and validation sets; initializing the multi-branch heterogeneous network model and the branch-level multi-scale loss constraint module; inputting the training set into the multi-branch heterogeneous network model and extracting training features through the first, second, and third branches respectively; and calculating the cross-entropy loss of the first branch, the center loss of the second branch, and the triplet loss of the third branch based on the training features. The cross-entropy loss and the triplet loss of the third branch are input into the branch-level multi-scale loss constraint module to calculate the intra-class similarity loss and inter-class distance loss; the total branch contrast loss is obtained based on the intra-class similarity loss and the inter-class distance loss; based on the total branch contrast loss, the branch parameters, each loss function parameter, and the parameters of the branch-level multi-scale loss constraint module of the multi-branch heterogeneous network model are iteratively updated using the backpropagation algorithm; during the iterative update process, the model recognition accuracy of the validation set is used; when the model recognition accuracy is greater than a preset accuracy threshold, training is stopped, and the trained multi-branch heterogeneous network model is obtained.

[0031] Radio frequency (RF) signal samples refer to raw RF signal data collected and labeled from various known drone models in actual or simulated environments. Each sample corresponds to a specific drone model and is used to train and validate the drone model recognition model. Training features refer to the high-dimensional feature vectors output by each branch after inputting the RF time-frequency maps from the training set into the first, second, and third branches of the multi-branch heterogeneous network model. These vectors serve as the basis for subsequent loss calculations. Intra-class similarity loss is a loss term calculated by the branch-level multi-scale loss constraint module based on the training features extracted from each branch. It measures the proximity of different samples of the same drone model in the feature space, aiming to reduce the feature distance between samples of the same class and improve intra-class consistency. Inter-class distance loss is a loss term calculated by the branch-level multi-scale loss constraint module based on the training features. It measures the separation of samples of different drone models in the feature space, aiming to increase the feature distance between samples of different classes and enhance inter-class separability. Total branch contrast loss is a comprehensive optimization objective obtained by fusing the intra-class similarity loss and inter-class distance loss with certain weights. It is used to uniformly guide the three heterogeneous branches in collaboratively learning a more discriminative joint feature representation. Branch parameters refer to the learnable parameters such as convolutional kernel weights and biases contained in the first, second, and third branches of a multi-branch heterogeneous network model. These parameters are continuously adjusted during training to optimize feature extraction capabilities. Loss function parameters refer to the learnable or adjustable parameters involved in cross-entropy loss, center loss, and triplet loss (e.g., the class center vector in center loss). These parameters are updated synchronously with the network parameters during training to adapt to task requirements. Aircraft type recognition accuracy refers to the proportion of samples on the validation set where the model correctly identifies all present aircraft types (i.e., the predicted label set completely matches the true label set) out of the total number of samples in the validation set for each radio frequency time-frequency image sample containing features of four or fewer drone models. This is used to evaluate the model's generalization performance in multi-aircraft mixed scenarios. The preset accuracy threshold is a performance standard set before model training. When the aircraft type recognition accuracy on the validation set reaches or exceeds this threshold, the model is considered converged, and training stops.

[0032] First, the drone identification system sequentially filters, performs time-frequency transformation, and standardizes raw radio frequency (RF) signal samples from various drone models to generate RF time-frequency maps in a uniform format. These maps are then randomly divided into training and validation sets to ensure consistent data distribution between training and evaluation. Second, the system initializes all parameters of the three branches of the multi-branch heterogeneous network model, along with their corresponding loss function modules and branch-level multi-scale loss constraint modules, establishing a starting point for subsequent training. Then, the RF time-frequency maps from the training set are simultaneously input into the three branches. Training features are extracted from these branches using a three-layer convolutional network in the first branch, a four-layer convolutional network in the second branch, and a five-layer triplet convolutional network in the third branch. Based on these features, the cross-entropy loss of the first branch (used for supervised classification), the center loss of the second branch (used to constrain the clustering of similar features), and the triplet loss of the third branch (used to increase the distance between different classes) are calculated.

[0033] Next, the system feeds these three losses into the branch-level multi-scale loss constraint module, fusing and calculating the intra-class similarity loss and inter-class distance loss, and then weighting them to obtain the total branch contrast loss, thereby comprehensively optimizing the collaborative discrimination ability of each branch. Subsequently, based on this total branch contrast loss, the system uses the backpropagation algorithm to synchronously update the branch parameters of the multi-branch heterogeneous network model, the learnable parameters in each loss function, and the weight parameters of the branch-level multi-scale loss constraint module, allowing the model to gradually converge. After each parameter update, the system uses a validation set to input the current model and calculate the model recognition accuracy to monitor generalization performance. Finally, when the model recognition accuracy reaches or exceeds a preset accuracy threshold, and remains above the preset accuracy threshold during multiple training sessions, the system determines that the model is sufficiently trained, stops iteration, and outputs the finally trained multi-branch heterogeneous network model.

[0034] Please refer to Figure 3 , Figure 3This is a schematic diagram of the CNN|Triplet-CRGM model provided in Embodiment 1 of the method for aircraft model identification based on radio frequency signals of multiple UAVs in this application. The model consists of three parallel branches, each corresponding to a different feature extraction path. The left side represents the "CNN / Triplet branch": the first branch has a three-layer convolutional structure with a cross-entropy loss function; the second branch has a four-layer convolutional structure with a center loss function; and the third branch has a five-layer convolutional triplet structure, optimized using a triplet loss function. The middle part is the "channel attention fusion module," which receives output features from the three branches, forms a joint feature representation through the concatenation of "three-branch features" and "enhanced features," and introduces a "three-branch attention mechanism" to weightedly fuse the features from each branch. The right side is the "local and global feature refinement module," which contains two parallel substructures: first, it processes CRIFA (Channel Residual Attention). After feature enhancement by the Fusion (channel residual attention fusion module), the upper part uses local refinement multi-branching and secondary enhancement through CRIFA to divide the features into three parts of detailed features. Generalized average pooling and one-dimensional convolution are then used to process the refined features. Below, a global feature aggregation sub-network extracts global features. The generated local and global features are then combined to generate channel attention weights and perform residual fusion. The two CRIFA outputs are then processed by generalized average pooling and fed into the "bifurcation contrast loss" module to calculate intra-class similarity and inter-class distance losses. Finally, all processed features are merged through a "feature concatenation" operation and fed into the "DNN classification layer," which consists of multiple fully connected layers to complete the final model classification task. The entire structure embodies a complete process of multi-branch heterogeneous design, cross-branch attention interaction, and local-global feature collaborative refinement.

[0035] Step S30: The first branch feature, the second branch feature, and the third branch feature are subjected to global feature encoding, branch attention generation, and feature weighting through the cross-branch attention interaction module to obtain attention-enhanced features; It should be noted that the cross-branch attention interaction module refers to a neural network structure used to model the interdependencies between features of different branches in a multi-branch heterogeneous network model. It generates attention weights reflecting the importance of each branch by calculating the global contextual information of each branch's features, and then dynamically weights and fuses the original branch features accordingly to achieve cross-branch information complementarity and feature enhancement. The attention-enhanced feature refers to the integrated feature representation output after inputting the first, second, and third branch features into the cross-branch attention interaction module, through global feature encoding, branch attention weight generation, and weighted fusion. This feature retains the specificity of each branch while strengthening the contribution of highly discriminative branches and suppressing redundant or noisy components, thereby improving the robustness and accuracy of subsequent model recognition.

[0036] As an example, the step of performing global feature encoding, branch attention generation, and feature weighting on the first branch feature, the second branch feature, and the third branch feature through a cross-branch attention interaction module to obtain attention-enhanced features includes: concatenating the first branch feature, the second branch feature, and the third branch feature along the channel dimension to obtain a global feature vector; compressing the global feature vector using a preset dimensionality reduction coefficient to obtain dimensionality-reduced global features; inputting the dimensionality-reduced global features into two fully connected layers, performing linear transformation and modified linear unit activation processing sequentially to obtain multi-channel intermediate features; performing Sigmoid activation processing on the multi-channel intermediate features to generate a first attention weight, a second attention weight, and a third attention weight; and weighting and fusing the first branch feature, the second branch feature, and the third branch feature according to the first attention weight, the second attention weight, and the third attention weight to obtain attention-enhanced features.

[0037] The global feature vector is a high-dimensional feature vector formed by directly concatenating the features of the first, second, and third branches along the channel dimension. It integrates all the original feature information extracted from the three branches and serves as the input basis for generating subsequent attention weights. The preset dimensionality reduction coefficient is a pre-set scaling factor used to control the size of the feature dimension after dimensionality reduction, balancing computational overhead and information retention. The dimensionality-reduced global feature is a low-dimensional feature representation obtained by passing the global feature vector through a linear transformation layer and compressing the number of channels according to the preset dimensionality reduction coefficient. This aims to reduce the complexity of subsequent processing and highlight key information. The multi-channel intermediate feature is a non-linear feature vector output by linearly mapping the dimensionality-reduced global feature through two fully connected layers and applying a modified linear unit activation function after the second layer. Its number of channels is the same as the original number of branches, and it is used to generate the attention weights corresponding to each branch. The first attention weight is a scalar weight value (ranging from 0 to 1) corresponding to the first branch after processing the multi-channel intermediate feature with a sigmoid activation function. It is used to measure the relative importance of the first branch feature in the fusion process. The second attention weight refers to the scalar weight value corresponding to the second branch after sigmoid activation, used to dynamically adjust the contribution of the second branch features to the final fusion result. The third attention weight refers to the scalar weight value corresponding to the third branch after sigmoid activation, used to adaptively enhance or suppress the role of the third branch features in the attention-enhanced features.

[0038] While multi-branch structures can extract differentiated features from RF time-frequency plots, they suffer from information isolation and insufficient utilization of complementarity between branches. By introducing a cross-branch attention interaction module and a branch-level multi-scale loss constraint module, adaptive fusion of branch information and multi-dimensional feature constraints are achieved.

[0039] The cross-branch attention interaction module takes the original time-frequency plot features of each branch as input and achieves bidirectional information complementarity between branches through global feature encoding, branch attention generation, and feature weighting enhancement. Let the original features of the three branches be... , , ( For the real number space, For branches k (Feature dimension), concatenate all branch features along the channel dimension to obtain the global feature vector: in, This is a splicing operation along the channel dimension. The dimension of the global features.

[0040] Then use global features FFor the input, an independent attention mechanism is designed for each branch, consisting of two fully connected layers, which generate the attention weights for the corresponding branches. To balance computational complexity and representational capability, a dimensionality reduction coefficient is introduced. r Compress the global feature dimension to Then map it back to the feature dimensions of the corresponding branch. (Based on branches) k Taking attention weight generation as an example: in, For branches k Attention weights; Use the Sigmoid activation function to ensure that the weight output range is [0,1]. To modify the activation function of the linear unit, nonlinearity is introduced; For the dimensionality reduction of the fully connected layer weight matrix ( ); For the weight matrix of the fully connected layer in higher dimensions; For the dimension-reduced fully connected layer bias vector; This is the bias vector for the upgraded fully connected layer.

[0041] The generated weight vector Time-frequency characteristics of the corresponding branch f k Element-wise multiplication yields attention-enhancing features: in, This is an element-wise multiplication operation.

[0042] Cross-branch attention only achieves feature fusion between branches, but does not adjust the distribution of the fused features. If the features extracted by different branches differ too much, it will affect the model's recognition performance to some extent. Therefore, a contrastive loss between branches is constructed to enhance the representation of information for each model. First, the time-frequency map features extracted by each branch are processed... L 2. Normalization to eliminate the impact of feature scale differences on distance calculation: in, for L 2-norm, This is a numerically stable term to avoid cases where the denominator is 0.

[0043] Calculate any branch pair Similarity matrix and construct a mask. (when When represents samples of the same type, (When representing outlier samples). The intra-class similarity loss is defined as: in, For the total number of samples of the same type, T This is an adjustment coefficient used to adjust the steepness of the similarity distribution (the smaller the value, the stronger the discrimination of similarity).

[0044] When enhancing the discriminative power of branch features for outlier samples from drones, cosine similarity is used as a basis for conversion. L 2 distance Construct inter-class masks ( , I (The identity matrix is ​​used to exclude samples from themselves). The inter-class distance loss function is: in, For the total number of out-of-class sample pairs, m A distance threshold is set to ensure that the characteristics of non-standard drones are always kept at a certain distance.

[0045] Finally, the average loss of all branch pairs is taken to obtain the total branch comparison loss: in, K For the number of branches, Summing all non-repeating branch pairs.

[0046] Backpropagation of the main branch comparison loss synchronously distributes the intra-class consistency and inter-class discriminative loss gradients to the parameters of each branch to achieve collaborative optimization and improve the discriminative power of the overall features of the UAV RF time-frequency map.

[0047] Step S40: The attention enhancement features are refined using the local and global feature refinement modules to obtain fused features; It should be noted that the local and global feature refinement module refers to a feature optimization structure composed of two cascaded sub-components: CRIFA and the Generalized Mean Pooling (GeM) module. CRIFA enhances local details and models inter-channel dependencies in the attention-enhanced features by introducing residual connections and channel attention mechanisms. GeM, on the other hand, performs learnable nonlinear pooling operations to adaptively aggregate spatial dimensional information and extract more discriminative global representations. Together, they achieve joint refinement of local fine structure and global semantic information. The fused feature refers to the final high-dimensional feature vector output after the attention-enhanced features have been processed sequentially by CRIFA and GeM in the local and global feature refinement modules. This feature retains the discriminative information after multi-branch interactions and further enhances the sensitivity to key local patterns and robust representation of the global structure through channel attention and learnable pooling, serving as input for subsequent model classification.

[0048] Understandably, the system first inputs attention-enhanced features into the local and global feature refinement modules, and then into the CRIFA module. A channel attention mechanism with residual connections is used to calculate and reweight each channel. Specifically, the input features are first subjected to global average pooling, then two fully connected layers and an activation function are used to generate channel weights. Finally, these weights are multiplied by the original features and residual paths are added to retain effective information and enhance discriminative channels. Next, the system passes the CRIFA output features to the GeM module. A learnable pooling exponent is used to perform a non-linear power-mean operation on the spatial dimension of each channel, followed by square root extraction. This adaptively focuses on more representative local regions, avoiding the limitations of traditional average or max pooling. Finally, the system outputs a fused feature that combines CRIFA-enhanced channel representation with GeM-refined global structure. This feature possesses both local detail sensitivity and global robustness, providing a more reliable basis for subsequent high-precision model classification.

[0049] Step S50: Input the fused features into the fully connected layer of the DNN and the Softmax activation function for classification to obtain the multi-drone model recognition results.

[0050] It should be noted that a fully connected layer in a DNN refers to a fully connected structure composed of multiple deep neural networks. Its input is fused features, and each neuron in each layer is connected to all neurons in the previous layer. Through a learnable weight matrix, high-dimensional features are non-linearly mapped and abstracted, progressively compressed and transformed into a low-dimensional discriminative representation suitable for classification. The multi-drone model recognition result refers to the probability distribution of each preset drone model category output by the Softmax activation function after processing the fused features through the fully connected DNN layer. The system selects the category with the highest probability as the final recognition decision, thereby achieving the identification of the specific drone model corresponding to the input radio frequency signal.

[0051] Understandably, the drone identification system first inputs the fused features into a fully connected DNN layer. This layer typically contains two to three fully connected neural networks. Each layer performs a weighted summation of all dimensions of the previous layer's output, adds a bias, and then processes the sum using a non-linear activation function (such as ReLU). This process gradually maps the high-dimensional fused features into low-dimensional, compact, and more discriminative category semantic vectors, aiming to eliminate redundant information and strengthen the decision boundaries related to the drone model label. Next, the system feeds the output of the last layer of the fully connected DNN into a Softmax activation function, calculating the exponentially normalized probability for each preset drone model category. This ensures that the sum of all category probabilities equals 1, thus obtaining a clear probability distribution that facilitates quantifying the model's confidence in each drone model. Finally, the system selects the category with the largest value in this probability distribution as the final judgment result, outputting a multi-drone model identification result to indicate the specific drone model most likely corresponding to the current input radio frequency signal.

[0052] The overall model composed of the multi-branch heterogeneous network model, the cross-branch attention interaction module, and the local and global feature refinement modules in this application is referred to as the CNN|Triplet-CRGM model.

[0053] This embodiment provides a method for aircraft type identification based on radio frequency signals of multiple UAVs. First, the UAV radio frequency signals are preprocessed to obtain a radio frequency time-frequency map, effectively suppressing out-of-band noise and retaining discriminative time-frequency fingerprints, providing high-quality input for subsequent identification. Second, a multi-branch heterogeneous network model is used to extract low-level, mid-level, and high-level features from the radio frequency time-frequency map using the first, second, and third branches respectively. These three branches are complementary in structure, capturing aircraft-specific information from different levels of abstraction. Then, a cross-branch attention interaction module is used to concatenate, reduce dimensionality, and perform nonlinear mapping on the features from the three branches, generating attention weights for each branch and weighted fusion to achieve dynamic complementarity and enhancement of important information between features. Next, a local and global feature refinement module refines the attention-enhanced features to obtain more discriminative and robust fused features. Finally, the fused features are fed into a fully connected layer of a DNN for high-level semantic mapping, and the aircraft type probability is output through a Softmax activation function to complete the final identification. This embodiment improves the accuracy of UAV radio frequency signal model identification and the ability to resist interference from complex electromagnetic environments in multi-aircraft coexistence scenarios through multi-level feature collaborative modeling and refined feature optimization.

[0054] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the aircraft type identification method for multiple UAV radio frequency signals according to this application. The local and global feature refinement module includes a channel residual attention fusion module and a generalized mean pooling module. Step S40 of the aircraft type identification method for multiple UAV radio frequency signals includes steps S41 to S45: Step S41: Reshape the attention enhancement features into a spatial feature map; Step S42: The spatial feature map is convolved and fused with residuals through the channel residual attention fusion module to obtain local enhanced features; Step S43: The local enhanced features are aggregated using the generalized mean pooling module to obtain global features; Step S44: The local enhancement feature is uniformly divided into multiple local features in the horizontal direction, and each local feature is enhanced and pooled twice to obtain multiple refined local features; Step S45: The global feature and multiple refined local features are concatenated along the channel dimension to obtain the fused feature.

[0055] It should be noted that the CRIFA module refers to a convolutional structure that combines channel attention mechanism with residual connection. It learns channel weights on the input features and adds the weighted features to the original input to enhance the response of important channels while retaining the original information, thereby improving the ability to express local details.

[0056] The GeM module refers to a non-linear pooling structure with learnable parameters. By introducing a trainable pooling exponent, it performs a power-mean operation on each location of the feature map and then takes the square root. This adaptively adjusts between average pooling and max pooling, effectively enhancing the ability to focus on salient regions and generating a more discriminative global representation. Spatial feature maps are transformed from one-dimensional or high-dimensional attention-enhancing features into two-dimensional feature maps with defined height, width, and channel dimensions through a reshape operation. This gives them spatial structure, facilitating subsequent spatial operations such as convolution and pooling. Local enhancement features are the feature maps output by the channel residual attention fusion module after performing convolution operations on the spatial feature maps and superimposing residual connections. This enhances the local responses of key channels while preserving the original spatial structure, improving sensitivity to subtle RF fingerprint patterns. Global features are vectors obtained by the generalized mean pooling module after non-linearly aggregating local enhancement features in the spatial dimension. These vectors do not contain spatial dimensions, retaining only channel information, and are used to characterize the global semantic properties of the entire input signal. Local features refer to multiple sub-region feature maps obtained by uniformly dividing the local enhanced features along the horizontal direction (i.e., the time or frequency axis) into several segments. Each segment corresponds to the structural information of the signal within a local time period or frequency band. Refining local features refers to the compact local representation obtained by further performing channel attention enhancement and pooling processing on each local feature. This not only preserves the uniqueness of the local time-frequency structure but also improves the discriminative ability through secondary enhancement.

[0057] As an example, the channel residual attention fusion module includes multiple sets of two-dimensional convolutional blocks, a global average pooling unit, a one-dimensional convolutional layer, a sigmoid activation function, and a residual fusion unit. The step of performing convolution and residual fusion on the spatial feature map through the channel residual attention fusion module to obtain local enhanced features includes: performing convolution, normalization, and ReLU activation on the spatial feature map through the multiple sets of two-dimensional convolutional blocks to obtain deep semantic features; performing global average pooling on the deep semantic features through the global average pooling unit to obtain channel global features; inputting the channel global features into the one-dimensional convolutional layer to obtain channel refined features; processing the channel refined features through the sigmoid activation function to generate channel attention weights corresponding to each channel; multiplying the channel attention weights element-wise with the deep semantic features to obtain weighted deep features; and performing residual fusion on the weighted deep features and the deep semantic features through the residual fusion unit to obtain local enhanced features.

[0058] It should be noted that multiple sets of 2D convolutional blocks refer to sub-networks composed of multiple sequentially connected 2D convolutional layers. Each set typically includes a convolution operation, batch normalization, and a ReLU activation function, used to extract spatial structure and semantic information from the spatial feature map layer by layer. The global average pooling unit calculates the average value of each channel in both the height and width spatial dimensions of the deep semantic features, compressing each channel into a scalar, thus generating a vector that retains only the channel dimension, used to represent the overall response strength of each channel. A one-dimensional convolutional layer refers to a one-dimensional convolution operation operating in the channel dimension. Its input is the channel's global features, and it models the dependencies between channels through learnable convolutional kernels to achieve non-linear interaction and refinement of channel information. The residual fusion unit is a structural module that adds weighted deep features and original deep semantic features element-wise. It is used to introduce an identity mapping path, prevent information loss, promote gradient flow, and improve model training stability and feature representation ability.

[0059] Deep semantic features refer to high-order feature maps output after multiple sets of two-dimensional convolutional blocks sequentially convolve, normalize, and activate the spatial feature map. They possess strong semantic abstraction capabilities and local detail representation. Channel global features are vectors obtained by performing global average pooling on deep semantic features using a global average pooling unit. Their length is equal to the number of channels, and each element represents the average activation level of the corresponding channel across the entire spatial range. Channel refinement features are new vectors output after processing channel global features through a one-dimensional convolutional layer. This convolutional operation performs nonlinear transformations on the original channel responses and models inter-channel interactions, providing a more refined basis for subsequent attention weight generation. Channel attention weights are vectors with values ​​between 0 and 1 obtained by inputting channel refinement features into a Sigmoid activation function. Each element corresponds to the importance weight of a channel, used to dynamically adjust the contribution of each channel in feature fusion. Weighted deep features are the result of multiplying the channel attention weights element-wise with the deep semantic features after expanding them along the channel dimension. This enhances important channels and suppresses redundant channels.

[0060] Understandably, the RF time-frequency map features output from each branch are concatenated to form a high-dimensional vector, and then converted into a spatial feature map using a reshape operation. F It provides structured input for parallel extraction of global and local features, and strengthens key features and suppresses redundant information through the channel attention mechanism of the CRIFA module.

[0061] Please refer to Figure 5 , Figure 5 This is a schematic diagram of the CRIFA module provided in Embodiment 2 of the method for aircraft model identification based on radio frequency signals of multiple UAVs in this application. The CRIFA module first converts the spatial feature map... F Deep semantic features are extracted using three sets of two-dimensional convolutional blocks. F 1, where "CBR" represents a combination operation of convolution, normalization, and ReLU activation, and "CB" unit represents convolution and normalization operations; then, through the "Avg" unit... F 1. Global Average Pooling (GAP) is used to capture channel information, resulting in a vector representing the channel mean. Then, a one-dimensional convolution with shared weights is introduced for refined learning. Finally, the weights between channels are obtained using the Sigmoid activation function. ; weight and F 1. After element-wise multiplication, and... F 1. Residual fusion is performed to obtain enhanced features. F 2. The specific process is expressed as follows: in, Indicates the kernel size as k The one-dimensional convolution process uses GAP, which is global average pooling. For element-wise multiplication, This is an element-wise addition.

[0062] Finally, the original input features F The result after processing with another set of convolutional blocks and F 2. Residual fusion is performed again, and the output features are enhanced by ReLU activation. F 3 (i.e., fusion feature) Its expression is: in, CB This refers to the convolutional block in the CRIFA module.

[0063] The enhanced features utilize both global feature branches and local refinement multi-branching to achieve parallel extraction of features at different granularities. The global feature branch directly applies GeM to the enhanced features to further extract global features with adaptive aggregation capabilities. The processing procedure of the generalized mean pooling module is represented as follows: in, H, W These represent the height and width of the local enhancement feature, respectively. p To adaptively adjust parameters, h This refers to the index of the feature map in the height direction. This refers to the index of the feature map along the width direction. This refers to the feature map (local enhancement feature) located at the th position. h line, number The pixel value at the column position (i.e., the feature intensity at that spatial location) is adaptively updated during the backpropagation process to strengthen the corresponding feature weights, balance the preservation of details and stability of global features, and finally output the global features. f g The overall structural information of the UAV's RF time-frequency map is preserved.

[0064] The local refinement multi-branch method evenly divides the enhanced features into three local features along the horizontal direction. Each local feature is then individually enhanced twice by the CRIFA module and processed by GeM pooling, and the local refinement features are output sequentially. f p1 , f p2 , f p3 This allows for the extraction of fine-grained details of the device model from different regions of the target, thus compensating for the lack of sensitivity of global features to local differences.

[0065] To fully utilize the complementarity of local and global features, global features are... fg With local features f p1 , f p2 , f p3 The data is concatenated and backpropagated using the same contrastive loss function to iteratively optimize the branch parameters, thereby enhancing the discriminative power of the fused features. Finally, the fused features are passed through a fully connected DNN layer combined with a Softmax activation function to output the probability distribution of drone signal categories, improving the performance of multi-drone identification and classification.

[0066] This embodiment first reshapes the attention-enhanced features, converting them into spatial feature maps with a clear spatial structure to facilitate subsequent convolution operations. Next, a channel residual attention fusion module performs multi-layer convolution, channel attention weighting, and residual fusion on this spatial feature map to obtain local enhanced features, effectively strengthening key channel responses while preserving original details. Then, a generalized mean pooling module performs nonlinear spatial aggregation on the local enhanced features to generate compact global features, improving robust representation of the overall signal pattern. Subsequently, the local enhanced features are uniformly divided into multiple local features along the horizontal direction, and each local feature undergoes further attention enhancement and pooling to obtain multiple refined local features, capturing discriminative details in different regions of the time-frequency image. Finally, the global features and all refined local features are concatenated along the channel dimension to form a fused feature, taking into account both global semantics and local structural information. This embodiment, through multi-level, multi-granular feature refinement, can improve the accuracy of UAV radio frequency signal model identification and the ability to resist complex electromagnetic interference in multi-UAV coexistence scenarios.

[0067] For example, to help understand the implementation process of the aircraft type identification method for multiple UAV radio frequency signals obtained by combining this embodiment with the above embodiment one, please refer to Figure 6 , Figure 6 A simplified flowchart of a method for identifying the type of multiple UAVs based on their radio frequency signals is provided, specifically: The diagram presents a top-down hierarchical structure. At the top is "Open Source and Real-World Data," representing the source of raw UAV RF signal data used for training and validation. Below this is the "Data Preprocessing" module, which performs preliminary processing such as filtering and noise reduction on the raw data. Next comes the "RF Time-Frequency Map" generation step, converting the preprocessed signal into a time-frequency domain representation. Following this is the "Multi-Branch Heterogeneous Feature Extraction" stage, divided into three parallel branches: Branch 1 uses a three-layer convolutional structure combined with a cross-entropy loss function to extract low-level local features; Branch 2 uses a four-layer convolutional structure combined with a center loss function to enhance intra-class cohesion; and Branch 3 uses a five-layer triplet structure combined with a triplet loss function to increase inter-class distance. The outputs of these three branches enter the "cross-branch attention interaction module," where feature fusion is achieved through global feature encoding and attention weight generation. This is followed by the "global feature encoding and attention weight generation" stage, which further optimizes the fused features. Next, the "feature weighting enhancement and bidirectional information complementarity" module enhances the expressive power of key features. Following this is the "local and global feature refinement" module, which enhances local features using the CRIFA module and extracts global features through GAP pooling. These two features are then "feature concatenated" to form a joint representation. Finally, the outputs are input into the "DNN fully connected recognition" module to complete the model classification task. The entire process culminates in "experimental verification" to evaluate model performance. This process embodies the end-to-end design from data acquisition to final recognition, covering key technical aspects such as signal processing, multi-branch feature extraction, attention fusion, and refined classification.

[0068] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the aircraft type identification method for multi-UAV radio frequency signals in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0069] Experimental verification: (1) Open source data simulation of multi-machine coexistence scenario: The DroneRFa dataset covers three open frequency bands: 915 MHz, 2.4 GHz, and 5.8 GHz. It includes 9 categories of urban outdoor drone flight control and image transmission signals and 15 categories of indoor drone signals. The outdoor data is divided into three categories according to distance: near (20~40 m), medium (40~80 m), and far (80~150 m).

[0070] Time-frequency analysis of UAV signals was performed on a Lenovo ThinkBook 15 G2 ITL device using the MATLAB 2021a platform. The experiment employed a 100 MHz sampling rate to meet the Nyquist criterion and UAV signal bandwidth requirements. A 16384-point Hamming window was selected to suppress spectral leakage. STFT was then applied to every 0.1 s segment of the data to obtain a continuous and stable time-frequency plot of the single UAV RF signal.

[0071] Without loss of generality, Phantom4 Pro, MATRICE 200, Air 2S, and MATRICE 600Pro drones from the dataset were selected and labeled as drones of categories A through D. By combining and superimposing signals from different single drones, a multi-drone coexistence scenario was simulated. Specifically, the sampling signals from categories A through D of the DroneRFa dataset in outdoor close-range scenarios (20-40m) at a frequency of 2.4-2.5 GHz were directly superimposed. It should be noted that the core of the experiment based on the open-source dataset was to verify the recognition effectiveness of the proposed model. To preserve the authenticity of the original data to the greatest extent possible and closely approximate actual signal characteristics, no additional adjustments were made to the signal amplitude in the dataset, nor were signal interference, distance attenuation, or other scenario-related effects introduced. In other words, complex environmental variables were excluded, focusing on verifying the model's ability to recognize and classify multi-drone coexistence signals. Time-frequency diagrams of some single-drone and multi-drone coexistence signals are shown.

[0072] (2) Model design verification: Seven sets of experiments were conducted, including inter-branch ablation, local and global feature refinement module ablation, to verify the rationality of the heterogeneous network design. Inter-branch ablation covered three-branch individual training and single-branch ablation scenarios. Individual branch training could evaluate the independent performance of each branch in extracting RF time-frequency map features and verify the effectiveness of each branch in representing device information. In single-branch ablation training, ablation of the first branch could verify the necessity of the three-layer convolutional structure for low-level feature extraction, ablation of the second branch could verify the contribution of the four-layer convolutional structure to integrating mid-level transition features, and ablation of the third branch could verify the value of the Triplet-CNN five-layer convolutional structure in extracting high-level features. Local and global feature refinement module ablation training could verify its ability to enhance local device detail perception and global semantic association, clarifying its important role in improving model recognition performance.

[0073] The ablation experiment was conducted on an Intel Core i5-1135G7 processor hardware platform and in the PyCharm software environment. The results are as follows: Figure 7 As shown: Depend on Figure 7 It is evident that in the inter-branch ablation experiments, the accuracy and convergence speed of each branch trained individually lagged behind the complete model; the ablation of any branch resulted in a decrease in recognition performance; and in the ablation scenarios of the local and global feature refinement module, removing this module also significantly reduced the model's accuracy. The comparative results demonstrate that heterogeneous networks, by fusing complementary features from each branch and utilizing the refinement module to further enhance the extraction of detailed information from the RF time-frequency map, effectively improve the overall recognition performance.

[0074] exist Figure 7In the training process, the loss value of the complete model decreased rapidly and tended to a low level in the early stage, while the loss converged relatively slowly for the models corresponding to each ablation experiment. Some models maintained a high loss value in the later stage, indicating that the proposed model can synergistically utilize the feature contributions of each branch and combine the enhancement effect of the refinement module to show the best training convergence performance.

[0075] (3) RF time-frequency diagram recognition test based on multi-machine coexistence: CNN, Triplet-CNN, Transformer, ResNet50, and MobileNetV2 were selected as comparison models to evaluate the performance advantages of heterogeneous networks in UAV RF time-frequency map recognition. All models used the Adam optimizer (with an initial learning rate of 0.0001, combined with a learning rate decay strategy to ensure training stability), and the batch size was set to 32 to balance training efficiency and generalization ability. The Softmax function was used for aircraft classification, and other key hyperparameters were dynamically iteratively optimized based on the gradient of the loss function using the backpropagation algorithm. To ensure the objectivity and fairness of the comparative evaluation, the core configurations of each model are as follows: CNN adopts a 5-layer convolutional and pooling structure, with parameters consistent with the heterogeneous network branches; Triplet-CNN reuses the third branch structure of the proposed model; Transformer adopts a 10×10 image block strategy, referring to the typical configuration of mainstream models, and introduces 128-dimensional position embedding and encoding into the input data, combined with multi-head attention and residual normalization processing; ResNet50 adopts a 50-layer residual network structure, with some parameters matching the feature extraction layers of heterogeneous networks; MobileNetV2 adopts a standard lightweight architecture with inverted residuals. Through homogenized basic parameter settings and differentiated architecture design, a multi-dimensional comparison of traditional, mainstream, and lightweight models with heterogeneous networks is achieved, comprehensively verifying the comprehensive advantages of the proposed model in terms of recognition accuracy, computational efficiency, etc.

[0076] During the training of single-unit models, the accuracy and loss function curves of each model for training the RF time-frequency map of a single type of UAV are as follows: Figure 8 As shown.

[0077] Depend on Figure 8 It is evident that the heterogeneous model achieves the fastest improvement in recognition accuracy, approaching a high level and remaining stable after 10 iterations; while the accuracy improvement of ResNet50, MobileNetV2, and Transformer models lags behind, and CNN and Triplet-CNN require more than 40 training iterations to achieve the best results. Figure 8In the early stages of training, the loss function of the proposed model showed a rapid downward trend, stabilizing at a low level later. In contrast, other comparative models exhibited varying degrees of oscillation or convergence lag, particularly with significant loss fluctuations in the early stages of training. In summary, the proposed model demonstrates significant advantages in recognition accuracy, convergence speed, and training stability, validating its effectiveness and superiority in UAV RF time-frequency image classification tasks.

[0078] Based on the model trained using open-source data, a recognition experiment was conducted on signals from multiple coexisting machines. The recognition accuracy is shown in Table 1. Table 1 (Accuracy of each model on the DroneRFa dataset multi-drone classification task) In scenarios where one to four types of drones coexist, the recognition accuracy of the designed model is 1, 0.90, 0.95 and 1, respectively, which is significantly better than the comparison models, indicating that the overall model has higher recognition accuracy and universality in multi-drone coexistence scenarios.

[0079] The model was trained based on real-world UAV datasets, and the accuracy and loss function curves are shown below. Figure 9 As shown in the figure, this diagram illustrates the model training performance parameters of a single UAV's RF time-frequency plot in a real flight scenario, consisting of two sub-plots on the left and right. The left subplot shows the accuracy change curves, with the horizontal axis representing the number of training iterations (from 0 to 120) and the vertical axis representing the accuracy (in percentage, ranging from 0.0% to 1.0%). The graph contains six different curves, representing the accuracy trends of six models—CNN, Triplet-CNN, Transformer, ResNet50, MobileNet, and Full Model—during training. The Full Model's curve reaches its highest point and stabilizes in the later stages. The right subplot shows the loss value change curves, with the horizontal axis also representing the number of training iterations (from 0 to 120) and the vertical axis representing the loss value (ranging from 0 to 8). This graph contains the same six curves as the left, reflecting the decreasing trend of the loss function value for each model during training. All curves gradually decrease with increasing training iterations, with the Full Model's loss value decreasing the fastest and ultimately reaching its lowest point. A small image is embedded in the upper right corner, showing magnified details of the early training phase (0 to 30 iterations), more clearly demonstrating the differences in convergence speed among the models in the early stages. Overall, this figure, by comparing the dynamic performance of different models in terms of accuracy and loss value, intuitively reflects that the Full Model has a faster convergence speed and higher final performance in single UAV recognition tasks.

[0080] The recognition accuracy of single to four types of UAVs coexisting in the actual flight test is shown in Table 2.

[0081] Table 2 (Accuracy of each model on the real-world dataset for multi-drone classification tasks) This application also provides a model identification device for multiple UAV radio frequency signals, please refer to... Figure 10 The aircraft type identification device for multiple UAV radio frequency signals includes: Preprocessing module 10 is used to preprocess the UAV radio frequency signal to obtain a radio frequency time-frequency diagram; Feature extraction module 20 is used to extract features from the radio frequency time-frequency map through a multi-branch heterogeneous network model to obtain first branch features, second branch features and third branch features; Feature enhancement module 30 is used to perform global feature encoding, branch attention generation and feature weighting on the first branch feature, the second branch feature and the third branch feature through cross-branch attention interaction module to obtain attention-enhanced features; The refinement module 40 is used to refine the attention enhancement features through local and global feature refinement modules to obtain fused features; The aircraft model recognition module 50 is used to input the fused features into the fully connected layer of the DNN and the Softmax activation function for classification, so as to obtain the multi-UAV model recognition results.

[0082] The aircraft model identification device for multiple UAV radio frequency signals provided in this application adopts the aircraft model identification method for multiple UAV radio frequency signals in the above embodiments, which can solve the technical problem of how to improve the accuracy of UAV radio frequency signal model identification and the ability to resist interference from complex electromagnetic environments in multi-UAV coexistence scenarios. Compared with the prior art, the beneficial effects of the aircraft model identification device for multiple UAV radio frequency signals provided in this application are the same as the beneficial effects of the aircraft model identification method for multiple UAV radio frequency signals provided in the above embodiments, and other technical features in the aircraft model identification device for multiple UAV radio frequency signals are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0083] This application provides a model identification device for multiple UAV radio frequency signals. The model identification device for multiple UAV radio frequency signals includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the model identification method for multiple UAV radio frequency signals in the above embodiment 1.

[0084] The following is for reference. Figure 11This document illustrates a structural schematic diagram of a device suitable for implementing the radio frequency signal identification of multiple drones according to embodiments of this application. The device for identifying the radio frequency signals of multiple drones in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 11 The illustrated device for identifying the type of drone based on radio frequency signals is merely an example and should not be construed as limiting the functionality or scope of use of the embodiments described in this application.

[0085] like Figure 11 As shown, the aircraft type identification device for multiple UAV radio frequency signals may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the aircraft type identification device for multiple UAV radio frequency signals. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the aircraft type identification device for multi-UAV radio frequency signals to exchange data with other devices wirelessly or via wired communication. Although the figure shows an aircraft type identification device for multi-UAV radio frequency signals with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0086] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0087] The aircraft model identification device for multiple UAV radio frequency signals provided in this application adopts the aircraft model identification method for multiple UAV radio frequency signals in the above embodiments, which can solve the technical problem of how to improve the accuracy of UAV radio frequency signal model identification and the ability to resist interference from complex electromagnetic environments in multi-UAV coexistence scenarios. Compared with the prior art, the beneficial effects of the aircraft model identification device for multiple UAV radio frequency signals provided in this application are the same as the beneficial effects of the aircraft model identification method for multiple UAV radio frequency signals provided in the above embodiments, and other technical features in the aircraft model identification device for multiple UAV radio frequency signals are the same as the features disclosed in the previous embodiment method, and will not be repeated here.

[0088] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0089] The above description is merely a specific 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.

[0090] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the aircraft type identification method for multi-UAV radio frequency signals in the above embodiments.

[0091] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0092] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a multi-UAV radio frequency signal identification device, the device performs the following actions: preprocesses the UAV radio frequency signals to obtain a radio frequency time-frequency map; extracts features from the radio frequency time-frequency map using a multi-branch heterogeneous network model to obtain first-branch features, second-branch features, and third-branch features; performs global feature encoding, branch attention generation, and feature weighting on the first-branch features, second-branch features, and third-branch features using a cross-branch attention interaction module to obtain attention-enhanced features; refines the attention-enhanced features using a local and global feature refinement module to obtain fused features; and inputs the fused features into a fully connected DNN layer and a Softmax activation function for classification to obtain multi-UAV model identification results.

[0093] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0094] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0095] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for identifying the type of UAVs based on radio frequency signals of multiple UAVs. This method can solve the technical problem of improving the accuracy of UAV radio frequency signal type identification and its ability to resist interference from complex electromagnetic environments in multi-UAV coexistence scenarios. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the UAV type identification method for radio frequency signals of multiple UAVs provided in the above embodiments, and will not be repeated here.

[0096] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for aircraft type identification based on radio frequency signals of multiple unmanned aerial vehicles (UAVs), characterized in that, The method includes: The radio frequency signal of the UAV is preprocessed to obtain the radio frequency time-frequency diagram; Feature extraction is performed on the radio frequency time-frequency map using a multi-branch heterogeneous network model to obtain the first branch feature, the second branch feature, and the third branch feature. The attention-enhanced features are obtained by performing global feature encoding, branch attention generation, and feature weighting on the first branch features, the second branch features, and the third branch features through the cross-branch attention interaction module. The attention enhancement features are refined using local and global feature refinement modules to obtain fused features; The fused features are input into a fully connected layer of a DNN and a Softmax activation function for classification, resulting in multi-drone model identification results.

2. The method as described in claim 1, characterized in that, The multi-branch heterogeneous network model includes a first branch, a second branch, and a third branch; The step of extracting features from the radio frequency time-frequency graph using a multi-branch heterogeneous network model to obtain first branch features, second branch features, and third branch features includes: The first branch features are obtained by extracting low-level features from the radio frequency time-frequency map through the three-layer convolutional structure of the first branch. The second branch features are obtained by extracting mid-level features from the radio frequency time-frequency map through a four-layer convolutional structure of the second branch. The radio frequency time-frequency map is subjected to high-level feature extraction through the five-layer triplet convolutional network of the third branch to obtain the third branch features.

3. The method as described in claim 1, characterized in that, The local and global feature refinement modules include a channel residual attention fusion module and a generalized mean pooling module; The step of refining the attention enhancement features through local and global feature refinement modules to obtain fused features includes: The attention-enhancing features are reshaped and converted into spatial feature maps; The spatial feature map is convolved and fused with residuals by the channel residual attention fusion module to obtain local enhanced features. The generalized mean pooling module is used to aggregate the local enhanced features to obtain global features. The local enhancement features are uniformly divided into multiple local features in the horizontal direction, and each local feature is enhanced and pooled twice to obtain multiple refined local features. The global feature and multiple refined local features are concatenated along the channel dimension to obtain the fused feature.

4. The method as described in claim 3, characterized in that, The channel residual attention fusion module includes multiple sets of two-dimensional convolutional blocks, a global average pooling unit, a one-dimensional convolutional layer, a Sigmoid activation function, and a residual fusion unit; The step of performing convolution and residual fusion on the spatial feature map through the channel residual attention fusion module to obtain local enhanced features includes: By performing convolution, normalization, and ReLU activation on the spatial feature map using the multiple sets of two-dimensional convolutional blocks, deep semantic features are obtained. The deep semantic features are subjected to global average pooling by the global average pooling unit to obtain the channel global features. The global features of the channels are input into the one-dimensional convolutional layer to obtain the channel refinement features; The channel refinement features are processed by the Sigmoid activation function to generate channel attention weights for each channel. The channel attention weights are multiplied element-wise by the deep semantic features to obtain weighted deep features; The weighted deep features and the deep semantic features are residually fused by the residual fusion unit to obtain local enhanced features.

5. The method as described in claim 3, characterized in that, The processing procedure of the generalized mean pooling module is as follows: in, H, W These represent the height and width of the local enhancement feature, respectively. p To adaptively adjust parameters, h This refers to the index of the feature map in the height direction. This refers to the index of the feature map along the width direction. This refers to the local enhancement feature located at the first... h line, number The pixel value at the column position.

6. The method as described in claim 1, characterized in that, The step of performing global feature encoding, branch attention generation, and feature weighting on the first branch features, the second branch features, and the third branch features through a cross-branch attention interaction module to obtain attention-enhanced features includes: The first branch feature, the second branch feature, and the third branch feature are concatenated along the channel dimension to obtain the global feature vector; The global feature vector is compressed in dimension by a preset dimensionality reduction coefficient to obtain dimensionality-reduced global features; The dimensionality-reduced global features are input into two fully connected layers, and linear transformation and modified linear unit activation are performed sequentially to obtain multi-channel intermediate features. The multi-channel intermediate features are subjected to Sigmoid activation to generate a first attention weight, a second attention weight, and a third attention weight. Based on the first attention weight, the second attention weight, and the third attention weight, the first branch feature, the second branch feature, and the third branch feature are weighted and fused to obtain the attention-enhanced feature; The formula for generating the first attention weight is as follows: in, k = 1, For branches k Attention weights It is the Sigmoid activation function. To modify the activation function of the linear unit, For the dimensionality reduction of the fully connected layer weight matrix ( ), For the weight matrix of the fully connected layer in higher dimensions; For the dimension reduction of the fully connected layer bias vector, This is the bias vector for the upgraded fully connected layer.

7. The method according to any one of claims 1 to 6, characterized in that, The multi-branch heterogeneous network model includes a first branch, a second branch, and a third branch. The first branch uses the cross-entropy loss function, the second branch uses the center loss function, and the third branch uses the triplet loss function. The training steps for the multi-branch heterogeneous network model include: Radio frequency signal samples from various types of drones are preprocessed and divided into training and validation sets. The multi-branch heterogeneous network model and the branch-level multi-scale loss constraint module are initialized. The training set is input into the multi-branch heterogeneous network model, and training features are extracted through the first branch, the second branch, and the third branch, respectively. The cross-entropy loss of the first branch, the center loss of the second branch, and the triplet loss of the third branch are calculated based on the training features. The cross-entropy loss, the center loss, and the triplet loss are input into the branch-level multi-scale loss constraint module to calculate the intra-class similarity loss and the inter-class distance loss. The total branch contrast loss is obtained based on the intra-class similarity loss and the inter-class distance loss; Based on the total branch contrast loss, the branch parameters, loss function parameters, and parameters of the branch-level multi-scale loss constraint module of the multi-branch heterogeneous network model are iteratively updated using the backpropagation algorithm. During the iterative update process, the computer model recognition accuracy is determined based on the validation set. When the model recognition accuracy is greater than the preset accuracy threshold, training stops, and a trained multi-branch heterogeneous network model is obtained.

8. A device for identifying the type of multiple unmanned aerial vehicles (UAVs) based on radio frequency signals, characterized in that, The device includes: The preprocessing module is used to preprocess the UAV radio frequency signal to obtain the radio frequency time-frequency diagram; The feature extraction module is used to extract features from the radio frequency time-frequency map through a multi-branch heterogeneous network model to obtain first branch features, second branch features and third branch features; The feature enhancement module is used to perform global feature encoding, branch attention generation, and feature weighting on the first branch feature, the second branch feature, and the third branch feature through the cross-branch attention interaction module to obtain attention-enhanced features; The refinement module is used to refine the attention enhancement features through local and global feature refinement modules to obtain fused features; The aircraft model recognition module is used to input the fused features into the fully connected layer of the DNN and the Softmax activation function for classification, so as to obtain the recognition results of multiple UAV models.

9. A device for identifying the type of multiple unmanned aerial vehicles (UAVs) based on radio frequency signals, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the aircraft type identification method for multiple UAV radio frequency signals as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the aircraft type identification method for radio frequency signals of multiple UAVs as described in any one of claims 1 to 7.

Images