Multi-uav radio frequency detection method based on physical prior and hyper-gaussian mask

Abstract

The application discloses a multi-unmanned aerial vehicle radio frequency detection method based on physical priori and hyper-Gaussian mask and belongs to the field of unmanned aerial vehicle signal identification.The core of the application is to introduce time-frequency physical characteristics into a deep learning detection framework.The application firstly converts an original radio frequency signal into a time-frequency spectrum diagram through a short-time Fourier transform;then, an asymmetric convolution kernel is used to adapt to the anisotropy of the signal, and adaptive initialization of multi-target query is carried out based on energy priori.Finally, a mask attention mechanism is constructed through a high-order spatial attenuation function, so that physical-level soft isolation of overlapping signals is realized.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) signal recognition, and more particularly to a multi-UAV radio frequency detection method based on physical priors and super Gaussian masks. Background Technology

[0002] With the rapid development of the low-altitude economy, drone swarm operations and dense multi-drone flights have become commonplace. However, this has also brought more serious public safety hazards such as illegal airspace intrusions, swarm interference, and malicious coordinated interference. Accurate identification and isolation in multi-drone concurrent scenarios has become a core and urgent need in the low-altitude security field. Among various sensing methods, radio frequency signal-based drone detection technology has become the mainstream solution in the industry due to its core advantages such as long operating range, strong adaptability to non-line-of-sight scenarios, and passive detection. However, existing technologies have the following limitations in multi-drone concurrent scenarios:

[0003] Existing RFID methods have limitations. Most existing RFID-specific deep learning methods (such as those fusing ZC sequence features or open set learning methods) treat the entire time spectrum as a whole for image-level classification, lacking pixel-level decoupling capabilities for multiple signal trajectories. When multiple drones operate simultaneously, broadband video transmission signals and discrete flight control signal trajectories frequently intersect. Existing models are prone to multi-drone signal confusion and blurred classification boundaries, making it difficult to accurately extract the signal trajectory and drone type for each drone.

[0004] Existing visual object detection models have limitations. Directly introducing general computer vision (CV) Transformer object detection models (such as DETR) to process radio frequency time spectrum results in a mismatch between architecture and physical characteristics.

[0005] 1) Signal adhesion problem: The global cross-attention mechanism of existing models is designed for isotropic natural images, while the temporal spectrum has strong anisotropy (time and frequency). When signals overlap densely, the unconstrained receptive field will cause the features of different UAVs to absorb each other, resulting in severe "signal adhesion" and mask breakage.

[0006] 2) Low query initialization efficiency: Existing models mostly use static embedding or blind dense grid sampling for target query initialization. In multi-UAV scenarios, they cannot adaptively locate the dispersed effective energy area, resulting in slow convergence and susceptibility to background noise interference.

[0007] In summary, existing technologies are not suitable for real-world low-altitude scenarios with multiple drones flying densely, and have the following shortcomings:

[0008] (1) Lack of multi-target concurrent detection capability: Most existing methods are optimized for single UAV scenarios. When the number of UAVs in the airspace increases and the signal overlap increases, the detection accuracy and recall rate drop sharply, and cannot support the concurrent and stable identification of 3 or more UAVs.

[0009] (2) Lack of overlapping signal decoupling capability: When multiple UAV signals overlap, the global attention mechanism of the existing method will cause serious “signal adhesion”, which cannot achieve physical-level soft isolation of different UAV signals, resulting in target ID switching, mask fusion, and inability to complete the separation of multiple UAV signals and UAV type identification;

[0010] (3) Poor adaptability of multiple signal types: In multi-drone scenarios, drones of different models and manufacturers simultaneously transmit narrowband frequency hopping flight control signals and broadband video stream signals. Existing methods cannot simultaneously capture the differentiated signal characteristics of multiple drones, which easily leads to missed detection of niche models and special signals.

[0011] (4) Insufficient cold start positioning capability for multiple targets: The blind query initialization mechanism of the traditional DETR cannot accurately locate multiple scattered signal areas at the same time in multi-UAV scenarios, which easily leads to multiple missed detections and repeated detections, slow convergence speed, and extremely poor engineering practicality.

[0012] Therefore, there is an urgent need for a UAV time-frequency signal detection method that natively supports concurrent identification of multiple UAVs, can achieve physical-level soft isolation of overlapping signals, adapts to multiple types of differentiated signals, and simultaneously achieves macroscopic positioning and fine extraction of single-UAV signal trajectories. Summary of the Invention

[0013] This invention provides a multi-UAV radio frequency detection method based on physical prior and super Gaussian mask, which natively supports concurrent identification of multiple UAVs, performs physical-level soft isolation for overlapping signals, adaptively adapts to multiple types of differentiated signals, and simultaneously achieves macroscopic positioning and fine extraction of single-UAV signal trajectories.

[0014] The technical solution adopted in this invention is as follows:

[0015] A multi-UAV radio frequency detection method based on physical priors and super-Gaussian masks includes the following steps:

[0016] Step 1: Acquire radio frequency signals and generate time-frequency characteristics

[0017] The mixed radio frequency signals in the target spatial domain are acquired and subjected to short-time Fourier transform to generate a two-dimensional time-frequency characteristic spectrum;

[0018] Step 2: Time-frequency asymmetric feature adaptation module extracts features

[0019] An initial feature map is obtained by multi-scale feature extraction of the two-dimensional time-frequency feature spectrum through a visual backbone network.

[0020] The initial feature map is input into the time-frequency asymmetric feature adaptation module, which performs multi-scale feature extraction based on the frequency branch of the vertical kernel and the time branch based on the horizontal kernel to obtain the target feature map.

[0021] Step 3: Generation of multi-objective queries based on physical priors

[0022] Calculate the spatial energy density map of the target feature map, and use the physical distribution characteristics to adaptively locate the active signal region to generate the target query sequence;

[0023] Step 4: Multi-machine signal isolation and decoding based on high-order spatial attenuation mask

[0024] For each target query in the query sequence, construct the corresponding spatial attention bias mask based on the relative coordinates of the target bounding box;

[0025] Among them, the spatial attention bias mask uses a high-order spatial decay function to assign high retention weights to features inside the target bounding box and exponentially suppress features outside the bounding box.

[0026] The spatial attention bias mask is subjected to overlapping interference features outside the target bounding box by a Transformer-based cross-attention computation unit to obtain the decoded features;

[0027] Step 5: Detect the output

[0028] For the decoded features, the multi-task prediction head outputs the UAV model, signal type, trajectory bounding box, and pixel-level segmentation mask.

[0029] Furthermore, in step 2, the time-frequency asymmetric feature adaptation module includes: a first spatial dimension convolutional layer, a frequency branch, a time branch, and a second spatial dimension convolutional layer;

[0030] The input to the first spatial dimension convolutional layer is the initial feature map, which is used to perform channel dimensionality reduction on the initial feature map, generate intermediate spatial feature maps, and send them to the frequency branch and time branch respectively.

[0031] The frequency branch adopts a first-depth separable convolutional structure, and its spatial convolution part uses multi-scale vertical kernels to extract the frequency dwell features and bandwidth of broadband video stream signals at different scales.

[0032] The temporal branch employs a second-depth separable convolutional structure, with its spatial convolutional part using multi-scale horizontal kernels to extract short-term temporal jump features and temporal coherence of narrowband flight control signals.

[0033] The input to the second spatial dimension convolutional layer is the element-wise addition and fusion result of the feature maps output by the frequency branch and the time branch, which is used for feature dimensionality enhancement and representation reconstruction.

[0034] The output feature map of the second spatial dimension convolutional layer is added to the initial feature map to obtain the output feature map of the time-frequency asymmetric feature adaptation module, which is the target feature map.

[0035] Furthermore, the first depthwise separable convolutional structure of the frequency branch includes vertical kernels of three scales: 5×1, 11×1, and 21×1.

[0036] Furthermore, the second depthwise separable convolutional structure in the temporal branch includes horizontal kernels of three scales: 1×3, 1×5, and 1×9.

[0037] It should be noted that the number of asymmetric convolution kernels (vertical / horizontal kernels) included in the frequency branch and the horizontal molecule does not need to be consistent. They correspond to independent feature extraction in the frequency dimension and time dimension, respectively. The number and size of the corresponding multi-scale convolutions can be configured according to the frequency band distribution and time domain jump characteristics of the target signal to achieve optimal representation of the two types of orthogonal physical features.

[0038] Furthermore, in step 3, the spatial energy density map of the target feature map is as follows:

[0039] Calculate the average absolute value of the target feature map along the channel dimension to obtain the initial energy density map;

[0040] The macroscopic density distribution in the frequency direction of the initial energy density map is calculated by sliding window, and the local activity distribution based on morphological dilation is calculated by fusing the two by taking the maximum value pixel by pixel to obtain the enhanced macroscopic energy density map. ;

[0041] calculate global mean with standard deviation Construct adaptive statistical thresholds ,in, This is the preset adjustment coefficient;

[0042] Will Below the threshold The region is set to zero, and the spatial energy density map of the target feature map is obtained.

[0043] Furthermore, in step 3, generating the target query sequence includes:

[0044] Non-maximum suppression (NMS) is used to extract spatially discrete candidate coordinates: a sliding window of preset size is used to traverse the spatial energy density map. If the energy value of the current pixel (i.e., the pixel value in the spatial energy density map) is the maximum value within the sliding window and is higher than the maximum value, then the candidate coordinates are selected. If so, then mark it as a local energy peak;

[0045] All extracted local energy peaks are sorted in descending order of energy intensity (i.e., the corresponding pixel value), and the top ones are selected. Spatial coordinates of points The target query sequence is generated by mapping the position of the query vector to a high-dimensional query vector through sine and cosine position encoding; where K is a preset value.

[0046] Furthermore, in step 4, the spatial attention bias mask is:

[0047] For each target query in the query sequence, its corresponding target bounding box is meshed to obtain the relative coordinates of any grid point of the target. ,in, The target number in the query sequence;

[0048] Construct the corresponding spatial attention bias mask based on the relative coordinates of the target bounding box;

[0049]

[0050] in, These are the pixel coordinates. Indicate target Spatial attention bias mask The mask value, The default penalty scaling factor. It is a higher-order exponent.

[0051] Furthermore, in step 4, the meshing process of the target bounding box is as follows: , ,in, The center coordinates of the target bounding box (i.e., the spatial coordinates of the corresponding target selected in step 3). This refers to the bounding box size.

[0052] Furthermore, in step 4, the cross-attention calculation unit queries the matrix during processing. It consists of content embeddings of multi-target query sequences superimposed with physical prior location encodings, with each vector corresponding to a query representation of a UAV target to be detected; key matrix The vector is composed of spatially flattened features superimposed with positional encoding from the target feature map, with each vector corresponding to a feature representation of a spatial location on the time-frequency map; value matrix The original spatially flattened features of the target feature map; spatial attention bias mask. Superimposed on the query matrix AND key matrix Based on the similarity calculation results, attention suppression of the region outside the bounding box is achieved.

[0053] Meanwhile, the present invention also provides a multi-UAV radio frequency identification device based on physical prior and super Gaussian mask, which includes:

[0054] Signal preprocessing module: used to acquire the mixed radio frequency signal in the target spatial domain and convert it into a two-dimensional time-frequency characteristic spectrum;

[0055] Asymmetric feature adaptation module: used to decouple and extract frequency-resident features and time-jump features from the two-dimensional time-frequency feature spectrum using asymmetric convolution kernels with different aspect ratios, so as to output an initial feature map;

[0056] Adaptive query generation module: used to calculate adaptive thresholds based on the spatial energy density of the initial feature map, and generate multi-target initial query sequences by extracting local energy peaks;

[0057] Mask isolation decoding module: Based on a high-order spatial decay function, it generates an independent spatial attention bias mask for each target query in the initial query sequence, and suppresses overlapping interference features outside the target bounding box in the Transformer-based cross-attention calculation unit to obtain the decoded features;

[0058] Multi-task prediction module: used to output the drone model, signal type, trajectory bounding box and pixel-level segmentation mask for each drone based on the decoded features.

[0059] The technical solution provided by this invention brings at least the following beneficial effects:

[0060] (1) Effectively suppressing signal adhesion phenomenon: This invention establishes a physical-level soft isolation barrier at the cross-attention level by introducing a dedicated bias mask based on a high-order spatial attenuation function. In scenarios where multiple machine signals frequently cross and overlap, it can effectively prevent the mutual absorption of different target features, and significantly reduce the target ID switching rate and mask breakage rate.

[0061] (2) Enhance the ability to detect multiple drones concurrently and improve generalization: The asymmetric feature adaptation mechanism designed in this invention fully considers the physical differences between the vertical direction (wideband spectrum) and the horizontal direction (time jump) of the time spectrum, and can simultaneously capture the features of drones with different communication protocols, which greatly improves the generalization ability of the model.

[0062] (3) Significantly accelerates model convergence and reduces false negative rate: This invention abandons the blind query initialization mechanism of traditional visual models and uses the physical prior of concentrated radio frequency signal energy to adaptively locate the query base point. This not only filters out a large amount of background interference, but also effectively avoids the problem of repeated detection and false negative in multi-target scenarios, thus improving the efficiency of engineering deployment. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 The flowchart shows a multi-UAV radio frequency detection method based on physical priors and super Gaussian masks.

[0065] Figure 2 A schematic diagram of the time-frequency asymmetric feature adaptation module;

[0066] Figure 3 This is a schematic diagram of the device implementation for a multi-UAV radio frequency detection method based on physical priors and super Gaussian masks.

[0067] Figure 4 This is a schematic diagram of the time-frequency graph output. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0069] This invention provides a multi-UAV radio frequency detection method based on physical priors and super-Gaussian masks. The core of this method lies in introducing time-frequency physical characteristics into a deep learning detection framework. First, the original radio frequency signal is converted into a time-spectrum image using a short-time Fourier transform. Then, an asymmetric convolution kernel is used to adapt the signal anisotropy, and adaptive initialization for multi-target queries is performed based on energy priors. Finally, a mask attention mechanism is constructed using a high-order spatial attenuation function to achieve physical-level soft isolation of overlapping signals.

[0070] In one embodiment, such as Figure 1The multi-UAV radio frequency detection method based on physical prior and super Gaussian mask provided in this embodiment specifically includes the following steps:

[0071] Step S1: Acquire radio frequency signals and generate time-frequency characteristics

[0072] A mixed radio frequency signal within the target airspace is acquired. This mixed signal contains continuous transmission variables from multiple concurrent UAVs and background noise. A short-time Fourier transform (STFT) is performed on the mixed radio frequency signal to generate a two-dimensional time-frequency feature spectrum, which is then used as input to the target detection model.

[0073] Step S2: Time-frequency asymmetric feature adaptation module extracts features

[0074] A pre-defined visual backbone network is used to extract multi-scale features from the two-dimensional time-frequency feature spectrum of the RF hybrid signal to obtain an initial feature map. The initial feature map is then input into the time-frequency asymmetric feature adaptation module for processing. Addressing the strong anisotropy of the UAV's RF time spectrum along the time and frequency axes, this module decouples the feature space to synchronously adapt to the differentiated features of the narrowband frequency-hopping flight control signal and the broadband video stream signal.

[0075] like Figure 3 As shown, the specific structure and extraction steps of the time-frequency asymmetric feature adaptation module are as follows:

[0076] (1) Initial spatial mapping: The initial feature map is taken as input and first passed through the first spatial dimension convolutional layer (such as a convolutional layer with a convolutional kernel of 1×1) for channel dimensionality reduction processing to generate an intermediate spatial feature map.

[0077] (2) Time-frequency orthogonal multi-scale decoupling: The spatial intermediate feature map is input in parallel into two independent asymmetric feature extraction branches, one is the frequency branch (vertical direction) and the other is the time branch (horizontal direction).

[0078] The frequency branch employs a set of multi-scale first asymmetric convolutional blocks to capture the frequency dwell characteristics and bandwidth of the broadband video stream signal at different scales. Specifically, the first asymmetric convolutional block uses a first depthwise separable convolutional structure, and its spatial convolution part uses a multi-scale first asymmetric convolutional kernel. The first asymmetric convolutional kernel is... (Vertical kernel) In this embodiment, the first asymmetric convolution kernel of the first depth separable convolution structure includes: 5×1, 11×1 and 21×1.

[0079] Temporal branch (horizontal direction): A set of multi-scale second asymmetric convolutional blocks is employed to capture the short-term temporal jump characteristics and temporal coherence of narrowband flight control signals. Specifically, the second asymmetric convolutional block uses a second depthwise separable convolutional structure, with its spatial convolutional portion employing multi-scale second asymmetric convolutional kernels. The second asymmetric convolutional kernel has a density of 1. (Horizontal kernel) In this embodiment, the second asymmetric convolution kernel of the second depth separable convolution structure includes 1×3, 1×5 and 1×9.

[0080] (3) Orthogonal feature fusion and reconstruction: The feature maps output by the frequency branch and the time branch are added and fused element by element, and then the fused features are sent to the second spatial dimension convolutional layer for feature dimensionality enhancement and representation reconstruction.

[0081] (4) Residual feature output: The residual connection mechanism is introduced to add the features of the reconstructed feature map to the initial feature map input by the module, and finally obtain the target feature map that has multi-scale frequency domain and time domain feature expression and effectively suppresses feature degradation, and output it to the next processing module.

[0082] Step S3: Generation of multi-objective queries based on physical priors

[0083] Calculate the spatial energy density map of the target feature map, adaptively locate active signal regions using physical distribution characteristics, and generate a target query sequence. The specific steps are as follows:

[0084] Step S301, Spatial Energy Map Construction: Calculate the average absolute value of the target feature map in the channel dimension to obtain the original energy density map. .

[0085] Step S302, Physical Prior Enhancement: To accommodate signals with different bandwidths, the macroscopic density distribution in the frequency direction (corresponding to the prior of the image transmission video signal) and the local activity distribution based on morphological dilation (corresponding to the prior of the narrowband control signal) are calculated by sliding window, and the two are fused to obtain the enhanced macroscopic energy density map. .

[0086] Step S303, Adaptive Background Removal: Calculation global mean with standard deviation Construct adaptive statistical thresholds: ,in, This is a preset adjustment coefficient. It will be lower than the threshold. The region is set to zero to achieve automatic shielding of background thermal noise.

[0087] Step S304, Local Energy Peak Extraction: Non-maximum suppression (NMS) is used to extract spatially discrete candidate coordinates. Specifically, a sliding window of a preset size (e.g., ...) is used. Traverse the energy map; if the energy value of the current pixel is the maximum value within the sliding window and is higher than... If a peak is detected, it is marked as a local energy peak. This mechanism ensures that the generated query points do not excessively accumulate within the same signal trajectory, achieving physical isolation of multiple targets.

[0088] Step S305, front Coordinate mapping: All extracted local energy peaks are sorted in descending order of energy intensity, and the top [values] are selected. (Preset value) spatial coordinates of points The spatial coordinates are mapped to high-dimensional query vectors using sine and cosine position encoding, serving as the initial input for the detection model for different UAV targets. In other words, each spatial coordinate is considered a target query. (Subscript) Number the target.

[0089] Step S4: Multi-machine signal isolation and decoding based on high-order spatial attenuation mask

[0090] For each target query in the query sequence, a corresponding spatial attention bias mask is independently constructed based on the predicted relative coordinates of the target bounding box. This mask employs a high-order spatial decay function, assigning high preservation weights to features inside the target bounding box and exponentially suppressing features outside the target bounding box. Wherein, the target... The center point of the target bounding box is Based on preset bounding box size This allows you to obtain the position and size of the target bounding box.

[0091] The bounding box of the target is meshed, targeting the first... Relative coordinates of any grid point of each target Its bias mask The form is expressed as:

[0092]

[0093] in, These are the pixel coordinates. Indicates the bias mask Each mask value, Image dimensions The default penalty scaling factor. For higher order exponents (preferably) (Super Gaussian order); the bias mask is independently injected into the cross-attention calculation unit of the Transformer to achieve parallel decoupling extraction of multi-machine overlapping signals and obtain decoding features.

[0094] The attention formula for the cross-attention calculation unit is as follows:

[0095]

[0096] in, These are the query matrix, key matrix, and value matrix, respectively. The query matrix is ​​the dimension of the key vector. It consists of content embedding of multi-target query sequences superimposed with physical prior location encoding. Each vector corresponds to a query representation of a UAV target to be detected. The specific formula is as follows:

[0097]

[0098] in Embed a matrix for learnable target query content, with each row corresponding to the semantic feature representation of a drone target to be detected; The physical prior query location encoding matrix generated in step S3 has each row corresponding to a sine and cosine location encoding of an energy peak coordinate, providing the target's initial spatial location prior for the query.

[0099] Key matrix It is composed of spatially flattened features superimposed with positional encoding from the target feature map. Each vector corresponds to the feature representation of a spatial location on the time-frequency map, and its specific formula is as follows:

[0100]

[0101] in The original feature matrix is ​​obtained by flattening the target feature map along the spatial dimension. Each row corresponds to the frequency domain-time domain fusion feature of a pixel position on the time-frequency map. This is a spatial location encoding matrix for the feature map, where each row corresponds to the sine and cosine position codes of a spatial pixel location, providing spatial location identifiers for the features. The query position encoding... Based on the physical prior of the energy peak of the input signal, an initial spatial anchor is provided for each target query; feature location encoding. A uniform spatial grid derived from the feature map provides a unified spatial index for all feature points. Together, they enable accurate attention matching of the target query in the time-frequency feature space.

[0102] Value matrix The original spatial flattening feature of the target feature map is given by the following formula:

[0103]

[0104] The value matrix directly uses the original features after flattening the target feature map, without superimposing positional encodings. This avoids positional information interfering with the purity of the feature semantics, ensuring the accuracy and effectiveness of the decoded features output after attention-weighted aggregation. Spatial attention bias mask. Superimposed on the query matrix AND key matrix Based on the similarity calculation results, attention suppression is achieved for the region outside the bounding box.

[0105] Step S5: The multi-task prediction head outputs the UAV model, signal type, trajectory bounding box, and pixel-level segmentation mask.

[0106] The multi-task prediction head includes: a model prediction head for predicting UAV models (classification task), a signal type prediction head for predicting signal types, a trajectory boundary prediction head for predicting trajectory bounding boxes (object detection task), and a segmentation mask prediction head for predicting pixel-level segmentation masks (semantic segmentation task).

[0107] After the output feature map of the cross-attention calculation unit is resized, it is fed into the corresponding task prediction head to obtain the UAV model, signal type, trajectory bounding box and pixel-level segmentation mask.

[0108] In the feature extraction stage, this invention utilizes a pre-trained visual backbone network (such as DINO) to extract multi-scale features F. To adapt to the different physical semantics of the time spectrum on the frequency axis and time axis, this example employs depth-separable asymmetric convolution in the feature adaptation module (FTAdapter) to address two major shortcomings of general symmetric convolution kernels in existing technologies: first, symmetric convolution kernels are designed for isotropic natural images, which is completely incompatible with the strong anisotropic physical characteristics of the time spectrum frequency-time axis; second, symmetric convolution cannot simultaneously decouple the orthogonal features of narrowband frequency-hopping flight control signals and broadband video stream signals of UAVs, which can easily lead to missed detections of niche models and special signals.

[0109] In a preferred embodiment, the specific dimensions of the asymmetric convolution kernels for depth-separable asymmetric convolution are set as follows: the frequency branch uses vertical kernels of 5×1, 11×1, and 21×1 to capture the frequency span of wideband signals; the time branch uses horizontal kernels of 1×3, 1×5, and 1×9 to capture the short-time time-hopping characteristics of the signal. Through this asymmetric design, the model can decouple frequency dwell and time hopping modes, significantly improving its ability to represent complex modulated signals.

[0110] During the query initialization phase, this invention locates active signal regions using a Physically Driven Query Generator (PDQG). Specifically, it first calculates a spatial energy density map. And using adaptive statistical thresholds Filter out background noise. Among them, adjustment coefficient The typical value range is [2.0, 5.0], and its preferred value is Based on Gaussian distribution The principle of selection is that this value can effectively cover most of the background thermal noise, ensuring that the initial query focuses only on the electromagnetic energy cluster with high confidence, thereby greatly accelerating the convergence speed of the model in multi-objective scenarios.

[0111] To address the "signal adhesion" problem when multiple UAV signals overlap, this invention introduces a Super Gaussian Mask Attention (SGMA) mechanism. Specifically, for each target query, this invention uses its predicted bounding box parameters (center...) With size ), construct a dedicated bias mask The bias mask is constructed using a high-order spatial decay function, and its preferred calculation formula is: ,in, , In the optimal solution of this example, the higher-order exponents... Set as Penalty scaling factor Set as That is, a fourth-order super-Gaussian decay function is used: .

[0112] The technical advantages of this design are as follows: the fourth-order function has a flatter top than the ordinary second-order Gaussian function, ensuring "lossless entry" of effective signal features within the bounding box; at the same time, its edge decay speed is extremely fast, enabling it to apply a huge negative bias to other overlapping signals in cross-attention calculation. Furthermore, the attention weights of externally overlapping targets will be exponentially suppressed to near zero, thus logically forming a physical-level hard isolation barrier and completely solving the problems of ID switching and mask fusion.

[0113] The method of this invention is based on the provided time-frequency domain asymmetric feature extraction mechanism, which is adapted to the differentiated signal characteristics of different drone models (such as different broadband and narrowband frequency hopping). By introducing physical energy prior, it realizes multi-target adaptive cold start positioning, improving query convergence efficiency and noise reduction capability. Based on the mask attention mechanism of high-order spatial attenuation function, it establishes physical-level soft isolation for different targets in the feature decoding stage, effectively suppressing the signal adhesion phenomenon in overlapping scenarios, and realizing accurate decoupling of multiple UAV signal trajectories and drone models.

[0114] Furthermore, embodiments of the present invention also provide a multi-UAV radio frequency identification device based on physical prior and super Gaussian mask, such as... Figure 3 As shown, the device includes:

[0115] Signal preprocessing module: used to acquire the mixed radio frequency signal in the target spatial domain and convert it into a two-dimensional time-frequency characteristic spectrum;

[0116] Asymmetric feature adaptation module: used to decouple and extract frequency-resident features and time-jump features from the time-frequency feature spectrum using asymmetric convolution kernels with different aspect ratios;

[0117] Adaptive query generation module: used to calculate adaptive thresholds based on the spatial energy density of feature maps, and generate multi-target initial query sequences by extracting local energy peaks;

[0118] Mask isolation decoding module: used to generate an independent attention bias mask for each target query based on a high-order spatial decay function, and to suppress overlapping interference features outside the bounding box in the cross-attention calculation;

[0119] Multi-task prediction module: used to output the classification results, macroscopic position and signal pixel-level segmentation mask of each UAV based on the decoded features.

[0120] This invention achieves accurate identification in multi-drone concurrent scenarios by physically modifying time-frequency features, query initialization, and the attention mechanism. Experiments show that even with three drones operating concurrently, the pixel-level Dice coefficient and model recognition rate remain above 80%. Figure 4 As shown, the decoding features are processed by a multi-task prediction head, and the aircraft type classification of multiple UAVs in the airspace, the signal type (image transmission signal or control signal, and its corresponding confidence probability) corresponding to each UAV signal trajectory, the bounding box coordinates of the signal trajectory, and the pixel-level segmentation mask used to characterize the signal trajectory of each UAV are output in parallel.

[0121] Those skilled in the art should understand that the above specific parameters (such as...) Convolution, adjustment coefficient 4th-order exponent, scaling factor The parameters (etc.) are only the optimal configuration for a specific dataset. In practical applications, depending on different monitoring environments and frequency resolution requirements, the above parameters can be replaced by equivalents or fine-tuned within a certain range, and all such changes should be covered within the protection scope of this invention.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0123] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A multi-UAV radio frequency detection method based on physical priors and super Gaussian masks, characterized in that, Includes the following steps: Step 1: Acquire radio frequency signals and generate time-frequency characteristics The mixed radio frequency signals in the target spatial domain are acquired and subjected to short-time Fourier transform to generate a two-dimensional time-frequency characteristic spectrum; Step 2: Time-frequency asymmetric feature adaptation module extracts features An initial feature map is obtained by performing multi-scale feature extraction on the two-dimensional time-frequency feature spectrum through a visual backbone network. The initial feature map is input into the time-frequency asymmetric feature adaptation module, which performs multi-scale feature extraction based on the frequency branch of the vertical kernel and the time branch based on the horizontal kernel to obtain the target feature map. Step 3: Generation of multi-objective queries based on physical priors Calculate the spatial energy density map of the target feature map, and use the physical distribution characteristics to adaptively locate the active signal region to generate the target query sequence; Step 4: Multi-machine signal isolation and decoding based on high-order spatial attenuation mask For each target query in the query sequence, construct a corresponding spatial attention bias mask based on the relative coordinates of the target bounding box; Among them, the spatial attention bias mask uses a high-order spatial decay function to assign high retention weights to features inside the target bounding box and exponentially suppress features outside the bounding box. The spatial attention bias mask is subjected to overlapping interference features outside the target bounding box by a Transformer-based cross-attention computation unit to obtain the decoded features; Step 5: Detect the output For the decoded features, the multi-task prediction head outputs the UAV model, signal type, trajectory bounding box, and pixel-level segmentation mask.

2. The method as described in claim 1, characterized in that, In step 2, the time-frequency asymmetric feature adaptation module includes: a first spatial dimension convolutional layer, a frequency branch, a temporal branch, and a second spatial dimension convolutional layer; The input to the first spatial dimension convolutional layer is the initial feature map, which is used to perform channel dimensionality reduction on the initial feature map, generate intermediate spatial feature maps, and send them to the frequency branch and time branch respectively. The frequency branch adopts a first-depth separable convolutional structure, and its spatial convolution part uses multi-scale vertical kernels to extract the frequency dwell features and bandwidth of broadband video stream signals at different scales. The temporal branch employs a second-depth separable convolutional structure, with its spatial convolutional part using multi-scale horizontal kernels to extract short-term temporal jump features and temporal coherence of narrowband flight control signals. The input to the second spatial dimension convolutional layer is the element-wise addition and fusion result of the feature maps output by the frequency branch and the time branch, which is used for feature dimensionality enhancement and representation reconstruction. The output feature map of the second spatial dimension convolutional layer is added to the initial feature map to obtain the output feature map of the time-frequency asymmetric feature adaptation module, which is the target feature map.

3. The method as described in claim 2, characterized in that, The first depthwise separable convolutional structure of the frequency branch includes three vertical kernels of three scales: 5×1, 11×1, and 21×1.

4. The method as described in claim 2, characterized in that, The second depthwise separable convolutional structure in the time branch consists of three horizontal kernels of three scales: 1×3, 1×5, and 1×9.

5. The method as described in claim 1, characterized in that, In step 3, the spatial energy density map of the target feature map is as follows: Calculate the average absolute value of the target feature map along the channel dimension to obtain the initial energy density map; The macroscopic density distribution in the frequency direction of the initial energy density map is calculated by sliding window, and the local activity distribution based on morphological dilation is calculated by fusing the two by taking the maximum value pixel by pixel to obtain the enhanced macroscopic energy density map. ; calculate global mean with standard deviation Construct adaptive statistical thresholds ,in, This is the preset adjustment coefficient; Will Below the threshold The region is set to zero, and the spatial energy density map of the target feature map is obtained.

6. The method as described in claim 5, characterized in that, Step 3, generating the target query sequence includes: Non-maximum suppression (NMS) is used to extract spatially discrete candidate coordinates: a sliding window of preset size is used to traverse the spatial energy density map. If the energy value of the current pixel is the maximum value within the sliding window and is higher than the maximum value, then the candidate coordinates are selected. If so, then mark it as a local energy peak; All extracted local energy peaks were sorted in descending order of energy intensity, and the top ones were selected. Spatial coordinates of points The target query sequence is generated by mapping the position of the query vector to a high-dimensional query vector through sine and cosine position encoding; where K is a preset value.

7. The method as described in claim 6, characterized in that, In step 4, the spatial attention bias mask is: For each target query in the query sequence, its corresponding target bounding box is meshed to obtain the relative coordinates of any grid point of the target. ,in, The target number in the query sequence; Construct the corresponding spatial attention bias mask based on the relative coordinates of the target bounding box; in, These are the pixel coordinates. Indicate target Spatial attention bias mask The mask value, The default penalty scaling factor. It is a higher-order exponent.

8. The method as described in claim 7, characterized in that, In step 4, the target bounding box is meshed as follows: , ,in, The center coordinates of the target bounding box This refers to the bounding box size.

9. The method as described in claim 7, characterized in that, In step 4, the cross-attention calculation unit queries the matrix during processing. It consists of content embeddings of multi-target query sequences superimposed with physical prior location encodings, with each vector corresponding to a query representation of a UAV target to be detected; key matrix The vector is composed of spatially flattened features superimposed with positional encoding from the target feature map, with each vector corresponding to a feature representation of a spatial location on the time-frequency map; value matrix The original spatial flattening features of the target feature map; Spatial attention bias mask Superimposed on the query matrix AND key matrix Based on the similarity calculation results, attention suppression of the region outside the bounding box is achieved.

10. A multi-UAV radio frequency identification device based on physical prior and super Gaussian mask, characterized in that, include: Signal preprocessing module: used to acquire the mixed radio frequency signal in the target spatial domain and convert it into a two-dimensional time-frequency characteristic spectrum; Asymmetric feature adaptation module: used to decouple and extract frequency-resident features and time-jump features from the two-dimensional time-frequency feature spectrum using asymmetric convolution kernels with different aspect ratios, so as to output an initial feature map; Adaptive query generation module: used to calculate adaptive thresholds based on the spatial energy density of the initial feature map, and generate multi-target initial query sequences by extracting local energy peaks; Mask isolation decoding module: Based on a high-order spatial decay function, it generates an independent spatial attention bias mask for each target query in the initial query sequence, and suppresses overlapping interference features outside the target bounding box in the Transformer-based cross-attention calculation unit to obtain the decoded features; Multi-task prediction module: used to output the drone model, signal type, trajectory bounding box and pixel-level segmentation mask for each drone based on the decoded features.

Images