An unmanned aerial vehicle radio frequency detection method based on anisotropy and time-frequency complementarity and related devices

Abstract

The application discloses a kind of based on anisotropy and time-frequency complementary unmanned aerial vehicle radio frequency detection method and related device, it is related to unmanned aerial vehicle radio frequency detection technical field, including obtaining space analog radio frequency signal, and baseband in-phase quadrature data sequence are obtained by down-conversion and analog-digital conversion, and three-channel anisotropic time-frequency tensor is obtained by short-time Fourier transform and physical threshold hard cut-off;It or intermediate layer feature is decoupled into time, frequency and identity reserved stream, respectively after executing one-dimensional convolution, carry out splicing, point-by-point convolution and residual connection, obtain anisotropic combination feature tensor;It is spliced into semantic and position stream with shallow high-resolution feature decoupling, and attention mask is obtained by specificity mapping and two-way orthogonal global pooling, cross-multiplication is added after fusion, and complementary feature tensor of broken reconnection is obtained;Based on its calculation energy weight multiplication obtains purification feature tensor;After shallow feature, complementary feature and purification feature multi-scale fusion, decoupling detection head is output early warning message.

Description

Technical Field

[0001] This application relates to the field of radio frequency detection technology for unmanned aerial vehicles (UAVs), and in particular to a UAV radio frequency detection method and related apparatus based on anisotropy and time-frequency complementarity. Background Technology

[0002] In UAV radio frequency detection within edge computing scenarios, the relevant technologies typically directly adopt general-purpose lightweight visual detection networks designed for natural images. These networks rely on two implicit conditions for deployment: first, the input image has an isotropic texture distribution, allowing for undifferentiated feature extraction using a square receptive field; second, the signal-to-noise ratio (SNR) is sufficiently high, and the signal exhibits a continuous and complete topological structure in the time-frequency domain, eliminating the need to consider long-range correlation reconstruction under fragmented conditions. However, in real-world low-altitude counter-terrorism operations, both conditions are simultaneously deviated from. To achieve ultra-real-time detection on portable hardware with strictly limited power consumption and heat dissipation, operators are forced to deploy general-purpose lightweight networks on edge computing platforms; simultaneously, to capture weak signals, the system must operate in complex electromagnetic environments with low SNR (0dB or even -20dB). This forced deployment directly led to two overlapping contradictions: First, although the general-purpose lightweight network has fewer than 2.5M parameters, its computational complexity still exceeds 6.0 GFLOPs, making it unable to consistently and stably output high frame rates of over 150 FPS on edge devices, resulting in the failure of its ultra-real-time capture capability for transient RF signals such as frequency hopping and bursts. Second, and more importantly, these networks use isotropic square convolutional kernels to process RF time-frequency maps, while the physical topology of RF signals on the time-frequency map exhibits significant anisotropy: for example, frequency hopping signals appear as long-duration, narrow-band "horizontal stripes," while burst broadband signals appear as instantaneous bursts with extremely wide "vertical stripes." When matching this asymmetric topology, the square receptive field generates invalid computation in a large number of background regions within the convolutional kernel, and it is difficult to establish long-range correlations across signal "breaks" caused by noise at low signal-to-noise ratios. The direct result of this is that when running the existing lightweight detection model on the edge computing platform, either frames are dropped due to excessive computational load, resulting in missed transient signals; or, under low signal-to-noise ratio, feature extraction misalignment and broken signals cannot be reconstructed, resulting in a large number of missed detections and full-screen false alarms, making it impossible for the detection system to complete the task of stable real-time early warning of unauthorized drones. Summary of the Invention

[0003] The purpose of this application is to provide a method and related device for radio frequency detection of unmanned aerial vehicles based on anisotropy and time-frequency complementarity. This method can significantly reduce the number of model parameters and computational complexity while greatly improving the detection recall rate and anti-false alarm capability of weak fracture signals in low signal-to-noise ratio environments.

[0004] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a UAV radio frequency detection method based on anisotropy and time-frequency complementarity, including: Acquire spatial analog radio frequency signals, and perform down-conversion and analog-to-digital conversion on the spatial analog radio frequency signals to obtain discrete baseband in-phase quadrature data sequences; A short-time Fourier transform is performed on the baseband in-phase orthogonal data sequence to generate a two-dimensional logarithmic power spectrum. Based on the preset physical absolute power lower threshold and physical absolute power upper threshold, the two-dimensional logarithmic power spectrum is dynamically range hard-truncated to generate a single-channel normalized tensor. The single-channel normalized tensor is then processed to obtain a three-channel anisotropic time-frequency tensor. First, the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network is used as the first input feature tensor. The first input feature tensor is then decoupled along the channel dimension in an asymmetric proportion into a time feature flow tensor, a frequency feature flow tensor, and an identity-preserving flow tensor. Next, one-dimensional convolutions with different directions are performed on the time feature flow tensor and the frequency feature flow tensor to obtain the one-dimensional convolution results. Finally, the one-dimensional convolution results are concatenated with the identity-preserving flow tensor along the channel dimension, and the concatenated tensor is subjected to pointwise convolution and residual connection to output the final anisotropic combined feature tensor. First, the final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted from the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor. The local feature tensor is then decoupled into a semantic initial flow tensor and a position initial flow tensor. Next, specific mapping is performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate a high-dimensional semantic tensor and a high-dimensional position tensor. Then, a dual-path orthogonal global pooling operation is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate a semantic attention mask tensor and a position attention mask tensor. Finally, the semantic attention mask tensor is multiplied element-wise with the high-dimensional position tensor to obtain the first multiplication result. The position attention mask tensor is multiplied element-wise with the high-dimensional semantic tensor to obtain the second multiplication result. Finally, the first multiplication result and the second multiplication result are added together and fused by pointwise convolution to output the final complementary feature tensor of break-reconnection. Based on the final complementary feature tensor, the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space is calculated. After nonlinear activation, energy attention weights are generated, and the energy attention weights are multiplied element by element with the final complementary feature tensor to output the purified feature tensor after background noise suppression. The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are fused at multiple scales and then input into the decoupled detection head to output pixel-level detection results. Based on the pixel-level detection results, early warning messages are output.

[0005] Optionally, the two-dimensional logarithmic power spectrum is dynamically range hard-truncated based on preset physical absolute power lower and upper threshold values ​​to generate a single-channel normalized tensor, specifically: ; in, This represents the single-channel normalized tensor generated after hard truncation; Represents a two-dimensional logarithmic power spectrum; This represents the lower limit threshold of physical absolute power. This represents the upper limit threshold of physical absolute power.

[0006] Optionally, the one-dimensional convolution result includes a horizontally mapped tensor and a vertically mapped tensor. First, the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network is used as the first input feature tensor. The first input feature tensor is then decoupled along the channel dimension in an asymmetric proportion into a time feature component tensor, a frequency feature component tensor, and an identity-preserving component tensor. Next, one-dimensional convolutions with different directions are performed on the time feature component tensor and the frequency feature component tensor to obtain the one-dimensional convolution result. Finally, the one-dimensional convolution result is concatenated with the identity-preserving component tensor along the channel dimension, and the concatenated tensor is subjected to pointwise convolution and residual connection to output the final anisotropic combined feature tensor, specifically including: Use the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network as the first input feature tensor; The first input feature tensor is decoupled according to the channel ratio of 1:1:2. The first quarter of the channels are extracted as the time feature flow tensor, the middle quarter of the channels are extracted as the frequency feature flow tensor, and the remaining half of the channels are used as the identity-preserving flow tensor. Perform a one-dimensional horizontal spatial convolution with a kernel of K×1 on the temporal feature flow tensor to generate a horizontal mapping tensor, and perform a one-dimensional vertical spatial convolution with a kernel of K×1 on the frequency feature flow tensor to generate a vertical mapping tensor; where K is an odd number of 5, 7 or 9. The horizontal mapping tensor, the vertical mapping tensor, and the identity-preserving flow tensor are concatenated along the channel dimension to obtain the concatenated tensor. Then, pointwise convolution and residual connection are performed on the concatenated tensor to output the final anisotropic combined feature tensor.

[0007] Optionally, the final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted in the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor, and the local feature tensor is decoupled into a semantic initial flow tensor and a positional initial flow tensor, specifically including: The final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted in the middle of the backbone network are concatenated through channel splicing to form a local feature tensor; The local feature tensor is decoupled along the channel dimension at a ratio of 1:3. The first quarter of the channels is extracted as the semantic initial flow tensor, and the remaining three-quarters of the channels are used as the position initial flow tensor.

[0008] Optionally, specific mappings are performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate a high-dimensional semantic tensor and a high-dimensional position tensor. Then, bi-path orthogonal global pooling operations are performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate a semantic attention mask tensor and a position attention mask tensor, specifically including: The semantic initial flow tensor is subjected to two consecutive spatial convolutions followed by channel expansion convolutions to generate a high-dimensional semantic tensor. In parallel, the position initial flow tensor is subjected to pointwise convolutions to generate a high-dimensional position tensor. Deep convolution is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate semantic local conditioning tensor and position local conditioning tensor; The semantic local conditioning tensor and the positional local conditioning tensor are subjected to dual-path spatial adaptive average pooling and element-wise multiplication, respectively, to obtain the semantic global interaction tensor and the positional global interaction tensor. Pointwise convolution and nonlinear activation function mapping are performed on the semantic global interaction tensor and the positional global interaction tensor respectively to generate the semantic attention mask tensor and the positional attention mask tensor.

[0009] Optionally, based on the final complementary feature tensor, the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space is calculated, specifically: ; in, This represents the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space. Represents the final complementary feature tensor; This indicates that each neuron deviates from the mean of its channel space. ; This represents the variance of each neuron's deviation from the spatial distribution of its channel. ; H represents the preset minimum regularization coefficient; H and W represent the height and width of the final complementary feature tensor, respectively. This indicates retrieving all channels; These represent the spatial index coordinates of the final complementary feature tensor on the height and width axes, respectively.

[0010] Optionally, the shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are fused at multiple scales and then input into the decoupled detection head to output pixel-level detection results. Based on the pixel-level detection results, an early warning message is output, specifically including: The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are input into the neck feature fusion network, which outputs a mesoscale feature tensor and a small-scale feature tensor. The neck feature fusion network is used for upsampling and splicing fusion through the feature pyramid. The mesoscale and small-scale feature tensors are input into the decoupled detection head, and class prediction and spatial coordinate regression are performed in parallel at each spatial grid point, outputting pixel-level detection results. The pixel-level detection results include pixel-level geometric regression bounding boxes, target presence confidence, and UAV model category probability distribution vectors. Based on preset short-time Fourier transform prior parameters, the pixel-level geometric regression bounding box is processed using the inverse geometric mapping function to generate a physical detection bounding box. The nonmaximum suppression algorithm is used to process the physical detection bounding box and the corresponding target existence confidence and UAV model category probability distribution vector to obtain the early warning message.

[0011] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the UAV radio frequency detection method based on anisotropy and time-frequency complementarity as described above.

[0012] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for radio frequency detection of unmanned aerial vehicles based on anisotropy and time-frequency complementarity.

[0013] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the anisotropic and time-frequency complementary UAV radio frequency detection method described above.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method and related apparatus for radio frequency detection of unmanned aerial vehicles (UAVs) based on anisotropy and time-frequency complementarity. By acquiring baseband signals and reconstructing three-channel anisotropic time-frequency tensors, and in particular by using a preset physical absolute power threshold to perform dynamic range hard truncation on the two-dimensional logarithmic power spectrum, the method solves the problem that traditional dynamic range normalization methods tend to amplify background thermal noise in complex electromagnetic environments. This achieves consistency and stability of background distribution in the time-spectrum graph under different signal-to-noise ratios, providing a clean and physically constrained network input for subsequent feature extraction, thereby enhancing the model's generalization ability in unknown and complex background noise environments.

[0015] By extracting anisotropic features based on asymmetric channel decoupling and directional one-dimensional convolution, the input feature tensor is decoupled into a time feature stream, a frequency feature stream, and an identity-preserving stream. Then, one-dimensional depth convolutions are performed laterally and vertically, respectively. This solves the computational redundancy problem in existing general-purpose visual detection networks that use isotropic square receptive fields to process radio frequency time-frequency maps. It achieves accurate matching of the lateral extension of frequency-hopping signals and the vertical extension of burst broadband signals. While significantly reducing the number of model parameters and the computational overhead of multiplication and addition, it preserves the key structural information of the physical topology of radio frequency signals. This solves the bottleneck between the energy efficiency limit of edge computing platforms and the computational overhead of mainstream networks, and realizes the ability to continuously and stably output high frame rate detection capabilities on low-power portable hardware.

[0016] By employing dual-path orthogonal global pooling operations—that is, performing one-dimensional adaptive average pooling along the time axis and frequency axis respectively and orthogonally fusing them to generate a global spatiotemporal attention mask—the problem of conventional spatial attention mechanisms easily generating erroneous responses to local high-energy noise patches under low signal-to-noise ratios is solved. At the same time, it breaks through the constraint of limited receptive fields of local convolutions, realizing cross-regional modeling of long-range dependencies of signals in the time and frequency domains, and providing accurate global context guidance for subsequent reconnection of broken signals.

[0017] By employing a time-frequency cross-complementary guidance mechanism, the local feature tensor formed by the convergence of deep features and shallow high-resolution features of the backbone network is decoupled into semantic flow and positional flow, generating high-dimensional semantic tensor and high-dimensional positional tensor respectively. Then, a bidirectional cross-attention mask is used to guide and weight the semantic and positional aspects, and finally, the complementary feature tensor of broken and reconnected signals is fused and output. This solves the problem of local breakage and fragmentation of weak UAV signals caused by noise masking in low signal-to-noise ratio environments, and realizes the synchronous reconstruction of broken signal fragments in terms of physical boundaries and category semantics, which greatly improves the detection recall rate.

[0018] By calculating energy weights based on the ratio of the square of the deviation of each neuron from the spatial mean of its channel to the spatial variance, and then reweighting the feature tensor after nonlinear activation, the problem of existing attention mechanisms relying on learnable parameters and being insufficient in suppressing smooth background noise at low signal-to-noise ratios is solved. This achieves the amplification of high-energy signal features and strong suppression of low-energy background noise without additional parameters, significantly reducing the false alarm rate in complex background noise environments.

[0019] By using multi-scale feature fusion and physical inverse geometric mapping, shallow high-resolution features, final complementary features, and purification features are fused through the neck network and input into the decoupled detection head, outputting pixel-level detection results. Based on short-time Fourier transform prior parameters, the pixel coordinates are inversely mapped to physically meaningful time-frequency detection bounding boxes, solving the problem of interpretability conversion of detection results from network output to actual electromagnetic space. This enables end-to-end early warning message generation that can be directly used by upper-layer business systems, completing the business closed loop of low-altitude UAV radio frequency detection. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 An application environment diagram for an anisotropic and time-frequency complementary UAV radio frequency detection method provided in Embodiment 1 of this application; Figure 2 A flowchart illustrating a UAV radio frequency detection method based on anisotropy and time-frequency complementarity provided in Embodiment 1 of this application; Figure 3 A schematic diagram of the overall architecture of a UAV radio frequency detection method based on anisotropy and time-frequency complementarity provided in Embodiment 1 of this application; Figure 4 This is a schematic diagram of an asymmetric parallel partial convolution (AP-Pzconv) feature extraction topology provided in Embodiment 1 of this application; Figure 5 A schematic diagram of a time-frequency feature cross-complementary guidance mechanism (TF-CCG) provided in Embodiment 1 of this application; Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment 4 of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1 The UAV radio frequency detection method based on anisotropy and time-frequency complementarity provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send spatial analog radio frequency signals to server 104. Server 104 acquires the spatial analog radio frequency signals, performs down-conversion and analog-to-digital conversion to obtain a baseband in-phase orthogonal data sequence, and performs short-time Fourier transform and physical threshold hard truncation to obtain a three-channel anisotropic time-frequency tensor. It decouples the time, frequency, and identity-preserving streams, performs one-dimensional convolution on each stream, and then concatenates, performs pointwise convolution, and performs residual connections to obtain anisotropic combined feature tensors. It concatenates and decouples these features with shallow high-resolution features to obtain semantic and positional streams, and performs specific mapping and dual-path orthogonal global pooling to obtain an attention mask. After cross-multiplication and addition, it obtains a broken and reconnected complementary feature tensor. Based on this, it calculates energy weights and multiplies them to obtain a cleaned feature tensor. After multi-scale fusion of shallow features, complementary features, and cleaned features, it outputs a warning message through a decoupled detection head. Server 104 can feed back the obtained warning message to terminal 102.

[0025] The terminal 102 can be, but is not limited to, various desktop computers, laptops, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0026] In one exemplary embodiment, such as Figures 2-3 As shown, a method for detecting unmanned aerial vehicles (UAVs) based on anisotropy and time-frequency complementarity is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 206. Wherein: Step 201: Acquire the spatial analog radio frequency signal, and perform down-conversion and analog-to-digital conversion on the spatial analog radio frequency signal to obtain a discrete baseband in-phase quadrature data sequence.

[0027] Step 202: Perform a short-time Fourier transform on the baseband in-phase orthogonal data sequence to generate a two-dimensional logarithmic power spectrum. Based on the preset physical absolute power lower threshold and physical absolute power upper threshold, perform dynamic range hard truncation on the two-dimensional logarithmic power spectrum to generate a single-channel normalized tensor. Then, process the single-channel normalized tensor to obtain a three-channel anisotropic time-frequency tensor.

[0028] Step 203: First, the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network is used as the first input feature tensor. The first input feature tensor is then decoupled along the channel dimension in an asymmetric proportion into a time feature flow tensor, a frequency feature flow tensor, and an identity-preserving flow tensor. Next, one-dimensional convolutions with different directions are performed on the time feature flow tensor and the frequency feature flow tensor to obtain the one-dimensional convolution results. Finally, the one-dimensional convolution results are concatenated with the identity-preserving flow tensor along the channel dimension, and pointwise convolutions and residual connections are performed on the concatenated tensor to output the final anisotropic combined feature tensor.

[0029] Step 204: First, the final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted from the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor. The local feature tensor is then decoupled into a semantic initial flow tensor and a position initial flow tensor. Next, specific mapping is performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate a high-dimensional semantic tensor and a high-dimensional position tensor. Then, a dual-path orthogonal global pooling operation is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate a semantic attention mask tensor and a position attention mask tensor. Finally, the semantic attention mask tensor and the high-dimensional position tensor are multiplied element-wise to obtain the first multiplication result. The position attention mask tensor and the high-dimensional semantic tensor are multiplied element-wise to obtain the second multiplication result. Finally, the first multiplication result and the second multiplication result are added together and fused by pointwise convolution to output the final complementary feature tensor of break-reconnection.

[0030] Step 205: Based on the final complementary feature tensor, calculate the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space. After nonlinear activation, generate energy attention weights and multiply the energy attention weights element by element with the final complementary feature tensor to output the purified feature tensor after background noise suppression.

[0031] Step 206: The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are fused through multi-scale features and then input into the decoupled detection head to output pixel-level detection results. Based on the pixel-level detection results, an early warning message is output.

[0032] Implementing steps 201 to 206 above will have the following beneficial effects: (1) By acquiring baseband signals and reconstructing three-channel anisotropic time-frequency tensors, especially by using a preset physical absolute power threshold to perform dynamic range hard truncation on the two-dimensional logarithmic power spectrum, the problem of the traditional dynamic range normalization method easily amplifying background thermal noise in complex electromagnetic environments is solved. This achieves consistency and stability of background distribution in the time-frequency spectrum under different signal-to-noise ratio conditions, providing a clean and physically constrained network input for subsequent feature extraction, thereby enhancing the model's generalization ability in unknown complex background noise environments.

[0033] (2) By extracting anisotropic features based on asymmetric channel decoupling and directional one-dimensional convolution, the input feature tensor is decoupled into time feature stream, frequency feature stream and identity preservation stream, and one-dimensional depth convolution is performed in the horizontal and vertical directions respectively. This solves the computational redundancy problem that exists when the existing general vision detection network uses isotropic square receptive fields to process radio frequency time-frequency maps. It achieves accurate matching of the horizontal extension of frequency hopping signals and the vertical extension of burst broadband signals. While significantly reducing the number of model parameters and the multiplication and addition computation overhead, it retains the key structural information of the physical topology of radio frequency signals. This solves the balance bottleneck between the energy efficiency ratio limitation of edge computing platforms and the computational overhead of mainstream networks, and realizes the ability to continuously and stably output high frame rate detection on low-power portable hardware.

[0034] (3) By performing a dual-path orthogonal global pooling operation, namely performing one-dimensional adaptive average pooling along the time axis and frequency axis respectively and performing orthogonal fusion to generate a global spatiotemporal attention mask, the problem that conventional spatial attention mechanisms are prone to making incorrect responses to local high-energy noise patches under low signal-to-noise ratio is solved. At the same time, it breaks through the constraint of limited receptive field of local convolution and realizes cross-regional modeling of long-range dependence of signal in the time and frequency domain, providing accurate global context guidance for subsequent reconnection of broken signals.

[0035] (4) Through the time-frequency cross-complementary guidance mechanism, the local feature tensor formed by the convergence of deep features and shallow high-resolution features of the backbone network is decoupled into semantic flow and position flow, respectively generating high-dimensional semantic tensor and high-dimensional position tensor. The semantic and position are mutually guided and weighted by the bidirectional cross-attention mask, and finally the complementary feature tensor of fracture reconnection is fused and output. This solves the problem of local fracture and fragmentation of weak UAV signals caused by noise masking in low signal-to-noise ratio environment, realizes the synchronous reconstruction of fracture signal fragments in physical boundary and category semantics, and greatly improves the detection recall rate.

[0036] (5) By calculating the energy weight based on the ratio of the square of the deviation of each neuron from the mean of the channel space to the spatial variance, and then re-weighting the feature tensor after nonlinear activation, the problem of existing attention mechanisms relying on learnable parameters and insufficient suppression of smooth background noise under low signal-to-noise ratio is solved. This achieves the amplification of high-energy signal features and strong suppression of low-energy background noise without additional parameters, significantly reducing the false alarm rate in complex background noise environments.

[0037] (6) Through multi-scale feature fusion and physical inverse geometric mapping, shallow high-resolution features, final complementary features and purification features are fused through the neck network and input into the decoupled detection head to output pixel-level detection results. Based on the short-time Fourier transform prior parameters, the pixel coordinates are inversely mapped to physically meaningful time-frequency detection bounding boxes, which solves the problem of interpretability conversion of detection results from network output to actual electromagnetic space. It realizes end-to-end early warning message generation that can be directly used by upper-layer business systems and completes the business closed loop of low-altitude UAV radio frequency detection.

[0038] In summary, this application effectively breaks the underlying design paradigm of traditional isotropic convolutional structures and conventional attention mechanisms by constructing a lightweight detection network with tightly coupled radio frequency physical priors. While significantly reducing the number of model parameters and computational complexity, it greatly improves the detection recall rate and anti-false alarm capability for weak fracture signals in low signal-to-noise ratio environments, making ultra-real-time UAV radio frequency detection for edge computing devices possible and providing a stable, reliable, and efficient technical solution for portable low-altitude countermeasure scenarios.

[0039] Step 201: Acquire the spatial analog radio frequency signal, and perform down-conversion and analog-to-digital conversion on the spatial analog radio frequency signal to obtain a discrete baseband in-phase quadrature data sequence.

[0040] Specifically, in the initial stage of the detection operation, the radio frequency front-end of the signal detection system monitors the target frequency band in real time in a complex electromagnetic environment, intercepting the simulated radio frequency signals of the UAV propagating in space (i.e., space simulated radio frequency signals). Subsequently, the computer equipment discretizes the space simulated radio frequency signals through conventional front-end signal processing mechanisms such as down-conversion and analog-to-digital conversion, obtaining the original UAV radio frequency baseband in-phase quadrature data sequence (i.e., discrete baseband in-phase quadrature data sequence). .

[0041] It should be noted that this embodiment does not impose strict limitations on the specific radio frequency acquisition hardware architecture and sampling rate. Any radio frequency front-end capable of outputting baseband in-phase quadrature sequences is applicable to this application.

[0042] Step 202 performs a short-time Fourier transform on the baseband in-phase orthogonal data sequence to generate a two-dimensional logarithmic power spectrum. Based on the preset physical absolute power lower threshold and physical absolute power upper threshold, the two-dimensional logarithmic power spectrum is dynamically truncated to generate a single-channel normalized tensor. The single-channel normalized tensor is then processed to obtain a three-channel anisotropic time-frequency tensor.

[0043] This step specifically includes the following sub-steps.

[0044] Sub-step 2.1: Time-frequency domain complex matrix mapping based on sliding time window.

[0045] The computer device retrieves the discrete baseband in-phase orthogonal data sequence obtained in step 201. The system pre-sets a time window function. Number of calculation points for short-time Fourier transform and the step size of the time window sliding The window function is used to process the baseband in-phase orthogonal data sequence. Perform a short-time Fourier transform (STFT) to generate a time-frequency complex matrix. Subsequently, the magnitude of the complex matrix is ​​extracted, and a minimal constant is introduced to prevent mathematical singularities. A base-10 logarithmic operation is then performed to transform it into a standard two-dimensional logarithmic power spectrum. The conversion calculation formula is as follows: ; in, The logarithmic power spectrum values ​​at time step λ and frequency point λ represent the calculated output, which constitute a two-dimensional spectrum containing frequency hopping and burst broadband physical topology features, and are the quantities to be determined. This represents a discrete baseband in-phase orthogonal data sequence acquired by a computer device, with its internal index... It achieves a precise mapping between global sampling points and local sliding window features, providing known quantities for system input; This represents the step size of the time window sliding, used to control the overlap rate and resolution of the time axis, and is a preset constant configured by the system. This represents the time window function (preferably a Hamming window or a Hanning window) applied to signal truncation to suppress spectral leakage; it is a preset constant. This represents the number of calculation points for the short-time Fourier transform, used to determine the physical resolution of the frequency axis, and is a preset constant. This represents the index of a discrete sampling point within a single time window, with a value range of [value missing]. , which is an intermediate unknown in the calculation; Represents the imaginary unit; This represents the minimum machine precision constant used to prevent mathematical singularities during logarithmic operations; it is a preset constant.

[0046] Sub-step 2.2: Dynamic range truncation based on a fixed physical threshold.

[0047] Computer equipment acquires the above two-dimensional logarithmic power spectrum Then, the system retrieves the preset physical absolute power lower threshold (representing the noise lower threshold, preferably -140dB) and physical absolute power upper threshold (representing the signal upper threshold, preferably -122dB to -100dB). Based on these fixed thresholds, the computer device performs a fixed dynamic range mapping calculation on the two-dimensional logarithmic power spectrum to generate a single-channel normalized tensor. The calculation formula is as follows: ; in, This represents the discrete time step index in the two-dimensional time-frequency space, and is a known quantity in the calculation; This represents the index of a discrete frequency point in the two-dimensional time-frequency space (i.e., the two-dimensional logarithmic power spectrum), which is a known quantity in the calculation; Indicates the preset physical absolute power lower limit threshold, subscript Characterizes the lower limit benchmark for blinding extremely low-energy background noise; Preset physical absolute power upper limit threshold, subscript Characterizes the upper limit reference used to lock the highest saturation brightness of the signal; This represents the single-channel normalized tensor generated after hard truncation, with its value range limited to [0, 1], and is an intermediate unknown quantity before entering the color mapping.

[0048] Sub-step 2.3: Pseudo-color visual mapping and bicubic interpolation.

[0049] To accommodate the input dimensions and multi-scale feature extraction requirements of the subsequent construction of a lightweight radio frequency feature parsing network, the computer device normalizes the single-channel normalized tensor. After discretization by multiplying by 255, the single-channel value is mapped to a pseudo-color RGB tensor containing three channels (red, green, and blue) using the system's built-in preset color mapping dictionary (preferably the Jet color space). Subsequently, the computer performs a flip-down geometric operation along the physical frequency axis on this RGB tensor to align it with the actual physical frequency rise and fall topology; and uses a bicubic interpolation algorithm to scale its spatial resolution to the network's preset standard input size (preferably 640×640), finally outputting a standard three-channel anisotropic time-frequency tensor that can be directly processed by deep learning frameworks.

[0050] Step 203 first uses the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network as the first input feature tensor, and decouples the first input feature tensor along the channel dimension in an asymmetric proportion into a time feature flow tensor, a frequency feature flow tensor, and an identity-preserving flow tensor; then performs one-dimensional convolutions with different directions on the time feature flow tensor and the frequency feature flow tensor respectively to obtain the one-dimensional convolution results; finally, the one-dimensional convolution results are concatenated with the identity-preserving flow tensor along the channel dimension, and pointwise convolution and residual connection are performed on the concatenated tensor to output the final anisotropic combined feature tensor.

[0051] This step is the asymmetric parallel partial convolution feature extraction process in this application, and its topological diagram is as follows: Figure 4 As shown. The asymmetric convolution kernel in this application is not designed for natural image textures in a conventional way, but is constructed based on the physical topological characteristics of the lateral frequency hopping continuity and longitudinal broadband burstiness exhibited by the UAV radio frequency signal in the time-spectrum image. Its receptive field direction and the physical extension direction of the signal form a one-to-one correspondence. This structure essentially constitutes a low-computational-cost directional decoupled convolution architecture for radio frequency anisotropic topology, specifically including the following sub-steps.

[0052] Sub-step 3.1: Decoupling of physical channels based on asymmetric ratio (Channel Split).

[0053] Computer devices can convert three-channel anisotropic time-frequency tensors or intermediate layer feature tensors in backbone networks. The first input feature tensor is used as the first input feature tensor. The first input feature tensor is decoupled according to the channel ratio of 1:1:2: the first 1 / 4 channels are extracted as the time feature flow tensor, the middle 1 / 4 channels are extracted as the frequency feature flow tensor, and the remaining 1 / 2 channels are used as the identity-preserving flow tensor.

[0054] Sub-step 3.2: Bidirectional independent feature mapping based on zero-padding spatial alignment and depth convolution.

[0055] For the three decoupled feature streams, the computer device performs completely independent three-way parallel processing. To ensure strict consistency of the output tensors of each branch in the spatial dimension to support subsequent residual reconstruction, an asymmetric edge zero-padding mechanism and a depthwise convolution strategy are introduced: Temporal coherent flow: The computer device's representation of the temporal characteristic flow sub-tensor Perform a one-dimensional lateral spatial convolution with a 1×K kernel, where K is an odd number (preferably 5, 7, or 9; in this embodiment, K=7). This operation uses channel-by-channel independent depthwise convolution and pads both sides of the tensor's width dimension (physical time axis) with (K-1) / 2 columns of zero values ​​to maintain spatial resolution, generating a lateral mapping tensor. .

[0056] Frequency-spreading flow: In parallel, the computer device processes the frequency-featured flow sub-tensor. Perform a one-dimensional vertical spatial convolution with a K×1 kernel, also using depthwise convolution, and pad the height dimension (physical frequency axis) of the tensor with (K-1) / 2 rows of zero values ​​above and below, generating a vertically mapped tensor. .

[0057] Identity-preserving current: for the identity-preserving current subtensor The computer device skips any spatial convolution operations and padding, and passes it directly as an identity mapping to minimize redundant computation and preserve the most original high-dimensional semantic features.

[0058] Sub-step 3.3: Cross-channel semantic fusion and residual reconstruction.

[0059] Due to the aforementioned precise asymmetric zero-filling strategy, the transverse mapping tensor... Vertical mapping tensor With identity-preserving flow carrier tensor Strictly aligned in the spatial dimension, the computer device concatenates the three components along the channel dimension to reconstruct a combined tensor with the channel count restored to full. Subsequently, point-wise convolution with a kernel size of 1×1 is used to perform cross-channel information interaction and global semantic fusion on the combined tensor. The fused output tensor is then pixel-wise added to the first input feature tensor (residual connection) to output the final anisotropic combined feature tensor of the current layer. The overall mathematical formula is as follows: ; ; in, This represents the intermediate layer feature tensor in the backbone network (i.e., the backbone feature extraction network), with dimension 1. , are known quantities; , Let C and C represent the time-dependent characteristic flux tensor and the frequency-dependent characteristic flux tensor, respectively, with a channel number of C / 4, which are intermediate unknowns. This represents the identity-preserving flux tensor with C / 2 channels, representing an intermediate unknown. The receptive field kernel length (preferably 7) represents the length of the receptive field of a one-dimensional asymmetric convolution, and is a preset constant for network configuration; : This represents the formula for calculating the number of zero-fill layers on one side to achieve residual alignment, and is a preset constant; These represent the depthwise convolution operators in the horizontal and vertical directions, respectively, and are preset constants; This represents a matrix operation operator that performs concatenation along the tensor channel dimension; it is a preset constant. This represents the three-dimensional combined tensor generated after the three streams are spliced ​​(i.e., the spliced ​​tensor), whose channel number is restored to C, which is an intermediate unknown; This represents a 1x1 pointwise convolution operator used for cross-channel semantic fusion, and is a preset constant. This represents the final anisotropic combination feature tensor output after the residuals are summed, and it is the quantity to be determined.

[0060] Step 204: First, the final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted in the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor. The local feature tensor is then decoupled into a semantic initial flow tensor and a position initial flow tensor. Next, specific mapping is performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate a high-dimensional semantic tensor and a high-dimensional position tensor. Then, a dual-path orthogonal global pooling operation is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate a semantic attention mask tensor and a position attention mask tensor. Finally, the semantic attention mask tensor is multiplied element-wise with the high-dimensional position tensor to obtain the first multiplication result. The position attention mask tensor is multiplied element-wise with the high-dimensional semantic tensor to obtain the second multiplication result. Finally, the first multiplication result and the second multiplication result are added together and fused by pointwise convolution to output the final complementary feature tensor of break-reconnection.

[0061] Before step 204, the following step is included: Step A, global spatiotemporal dependency feature modeling based on anisotropic fringe pooling: This step is not strictly stacked sequentially with the preceding steps, but rather aims to construct a general anisotropic fringe pooling mask generation mechanism (abstracted as a general operator) in the network. For any input feature tensor introduced into this mechanism (this input term serves as a mathematical parameter, generally referring to any intermediate layer tensor in the network that requires long-range spatiotemporal context dependency modeling), this step extracts its frequency energy profile and temporal persistence envelope through dual-path orthogonal fringe pooling, and reconstructs the global spatiotemporal attention mask using a tensor broadcasting mechanism. This mechanism is independent of a specific hierarchical order and can not only be applied to the global reweighting enhancement of midstream features in the backbone network, but also be configured to be independently instantiated and called by multiple heterogeneous feature flow branches in subsequent step 204.

[0062] Sub-step A.1 Local context preconditioning based on depthwise convolution.

[0063] To aggregate local neighborhood features before performing wide-range pooling, the computer device first processes the input tensor. A 3x3 kernel with 1 padding is applied. This operation extracts local texture features from each channel independently, generating a local conditioning tensor, while maintaining the spatial dimension and number of channels. .

[0064] Sub-step A.2 Two-way orthogonal strip pooling along the physical coordinate axes.

[0065] Computer equipment for local conditioning tensors Decoupled dual-path spatial adaptive average pooling is implemented to compress it along the time and frequency dimensions.

[0066] Frequency profile extraction branch: The computer performs adaptive average pooling along the width dimension of the feature tensor (corresponding to the physical time axis), compressing it to 1 while retaining the complete height dimension (corresponding to the physical frequency axis). This operation will... The tensor mapping is column feature vectors It accurately characterizes the global energy distribution of the signal in different frequency bands.

[0067] Temporal envelope extraction branch: In parallel, the computer performs adaptive average pooling along the height dimension (corresponding to the physical frequency axis) of the feature tensor, compressing it to 1 while retaining the full width dimension (corresponding to the physical time axis). This operation maps it to... The row eigenvectors accurately characterize the temporal envelope of the signal throughout the entire observation period.

[0068] Sub-step A.3 Time-frequency orthogonal crossover and mask generation based on tensor broadcasting mechanism.

[0069] To establish global interaction across the time and frequency dimensions, the computer device utilizes the Tensor Broadcasting mechanism of the deep learning framework to transmit the extracted column feature vectors. With row feature vectors Perform element-wise multiplication (Hadamard Product). This orthogonal crossover operation not only mathematically reconstructs a spatial dimension of... The global interaction tensor also ensures that any pixel within the tensor simultaneously contains global statistical information about its frequency band and time window.

[0070] Subsequently, the computer device uses pointwise convolutions with a kernel size of 1x1 to fuse cross-channel information in the cross-process results, and then uses a Sigmoid activation function to nonlinearly map its values ​​to the (0,1) interval, finally outputting a two-dimensional global attention mask tensor of spatial channels. Its core mathematical formula is as follows: ; ; ; ; in, This represents the initial feature tensor input to the stripe pooling module, which is a known quantity in the current step; This represents the intermediate feature tensor after 3x3 depthwise convolution preconditioning, with spatial dimensions of... , which is an intermediate unknown; This represents the operator that performs one-dimensional adaptive average pooling along the tensor width dimension (time axis), and is a preset constant; This represents the operator that performs one-dimensional adaptive average pooling along the tensor height dimension (frequency axis), and is a preset constant; Let the column eigenvector tensor representing the frequency-energy profile have spatial dimensions of . , which is an intermediate unknown in the calculation; Let the row eigenvector tensor representing the temporal continuum envelope have spatial dimensions of . , which is an intermediate unknown in the calculation; This represents the element-wise product operator (Hadamard Product) based on the tensor broadcast mechanism, used to implement the orthogonal space expansion of row and column vectors, and is a preset constant; This represents the unbiased convolution operator with a kernel size of 1x1 used to generate the attention mask, which is a preset constant; This represents the Sigmoid nonlinear activation function operator, used to constrain eigenvalues ​​to (0,1), which is a preset constant; This represents the final generated global spatiotemporal attention mask tensor, whose spatial dimensions are restored to . , is used to guide the subsequent feature representation of re-weighting, and is the variable to be determined in this embodiment.

[0071] Furthermore, when this mechanism performs global augmentation as an independent operator in the backbone network, the computer device will generate a global spatiotemporal attention mask tensor. With local conditioning tensor - Perform element-wise multiplication along the channel dimension to complete the global spatiotemporal reweighting of features, and output the enhanced feature tensor. : ; in, This represents the enhanced feature tensor after spatial channel dual-dimensional attention reweighting, which eliminates invalid background redundancy and strengthens long-range spatiotemporal dependence. It is the final output of this operator.

[0072] Step 204 is the time-frequency cross-complementary guidance mechanism in this application, and its schematic diagram is shown below. Figure 5 As shown. Specifically, it includes the following sub-steps.

[0073] Sub-step 4.1: Hard decoupling of semantics and location.

[0074] The computer device concatenates the final anisotropic combined feature tensor with shallow high-resolution feature tensors extracted from the backbone network via channel splicing to form a local feature tensor. This local feature tensor... Decoupled along the channel dimension at a ratio of 1:3: Extract the first 1 / 4 of the channels as the semantic initial flow tensor Use the remaining 3 / 4 channels as the initial flow tensor for the location. .

[0075] Sub-step 4.2: Specific receptive field mapping and channel alignment of heterogeneous branches.

[0076] The computer device performs completely independent specific convolution operations on the decoupled dual branches and forces the channel dimensions to be aligned to the original dimension C at the output to support subsequent matrix multiplication and residual addition: Computer device for the semantic initial stream tensor Two consecutive spatial convolutions with a kernel size of 3×3, padding of 1, and stride of 1 are performed to expand the receptive field and deeply mine the implicit topological texture of the fracture signal. Then, a channel-expanding convolution with a kernel size of 1×1 is applied to stretch the number of channels from 1 / 4 of the total number of channels to the full number of channels, generating a high-dimensional semantic tensor. .

[0077] In parallel, the computer device initializes the flow tensor at the location. Perform a pointwise convolution with a kernel size of 1×1, completely skipping neighborhood aggregation in the spatial dimension to maintain the original temporal and frequency coordinates with the highest fidelity, while mapping the number of channels from 3 / 4 to the full number of channels, generating a high-dimensional location tensor. .

[0078] Sub-step 4.3: Generate cross-attention masks based on dual-path orthogonal global pooling.

[0079] The computer devices respectively process the high-dimensional semantic tensor and the high-dimensional position tensor As input (i.e., substituted into step A) Perform bi-path orthogonal global pooling to generate semantic attention mask tensors. and position attention mask tensor The dual-path orthogonal global pooling operation specifically includes: first, performing a depthwise convolution with a kernel size of 3×3 and an edge padding of 1 on the input tensor to generate a local conditioning tensor; then, performing one-dimensional adaptive average pooling along the width dimension (time axis) of the local conditioning tensor to obtain frequency profile column feature vectors, and simultaneously performing one-dimensional adaptive average pooling along the height dimension (frequency axis) to obtain temporal envelope row feature vectors; using a tensor broadcasting mechanism, performing element-wise multiplication of the column feature vectors and the row feature vectors to generate a global interaction tensor; finally, performing a pointwise convolution with a kernel size of 1×1 on the global interaction tensor, and mapping it with a Sigmoid nonlinear activation function to generate an attention mask tensor.

[0080] Following the above operations, a positional attention mask tensor is generated using a high-dimensional position tensor as input, and a semantic attention mask tensor is generated using a high-dimensional semantic tensor as input.

[0081] Sub-step 4.4: Residual aggregation and final projection of complementary features.

[0082] The computer device performs element-wise multiplication of the semantic attention mask tensor and the high-dimensional position tensor to obtain the first multiplication result. (The semantically guided position tensor); perform element-wise multiplication of the position attention mask tensor with the high-dimensional semantic tensor to obtain the second multiplication result. (Location-guided semantic tensor). Since the first and second multiplication results are strictly aligned in both spatial and channel dimensions, the computer adds them element-wise to achieve deep complementary aggregation of physical properties. Finally, cross-channel information smoothing is performed on the aggregation result using a 1×1 projective convolution, outputting the final complementary feature tensor with broken and reconnected features. .

[0083] ; ; ; ; ; in, This represents a local input feature tensor formed by the aggregation of different levels, and is a known input quantity; , : These represent the semantic initial flow tensor and the position initial flow tensor decoupled at ratios of 1 / 4 and 3 / 4, respectively, and are known intermediate quantities in the computation; : Indicates that the core size is Furthermore, the spatial convolution operator that maintains the feature map size is used for deep texture mining and is a preset constant; : These represent the pointwise convolution operators used for channel expansion, channel alignment, and final feature projection, respectively, and are preset constants; : These represent the high-dimensional semantic tensor and high-dimensional position tensor after receptive field expansion and channel alignment, respectively, and are intermediate unknowns; : respectively represent and As input, the positional attention mask tensor and semantic attention mask tensor, rigorously calculated through the mathematical process of step four (striated pooling) mentioned above, both have the same dimension. , where is a known guiding quantity; : Represents an element-wise product operator based on the tensor broadcast mechanism, which is a preset constant; : These represent the semantic tensor and position tensor after cross-mask reweighting and guidance, respectively, and are intermediate unknowns; : Represents the final complementary feature tensor output after performing cross-branch complementary aggregation and projection, which is the quantity to be determined.

[0084] Step 205 calculates the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space based on the final complementary feature tensor. After nonlinear activation, energy attention weights are generated and multiplied element-wise with the final complementary feature tensor to output the purified feature tensor after background noise suppression.

[0085] This step is the parameterless energy attention mechanism in this application, used for adaptive hard suppression of strong background noise. The specific process is as follows: Sub-step 5.1 Statistics on the mean and variance of neurons in the spatial dimension.

[0086] To establish a local neighborhood reference for each feature channel, the computer device first moves along the complementary feature tensor. Calculate the spatial mean of all neurons within each channel using the height (physical frequency axis) and width (physical time axis) dimensions. Then, the square of the difference between the activation value of each neuron and the mean value of that channel is calculated and divided by the total number of pixels minus one (i.e., ...). Thus, the spatial variance characterizing the background energy fluctuations of the channel is obtained. .

[0087] Sub-step 5.2 Evaluation of three-dimensional energy closed-form solutions based on regularization.

[0088] After obtaining the channel statistics, the computer device directly calls the closed-form solution formula based on the neuron energy minimization problem to evaluate... For each target neuron in the dataset, calculate the squared deviation of that neuron from the background mean and divide it by the spatial variance and the system's preset regularization coefficient. (Preferred) The sum of these terms is four times the sum of the given terms, and finally a fixed bias term of 0.5 is added. This introduces a very small regularization coefficient. The aim is to prevent division-to-zero singularities and to apply smooth constraints to energy extrema. The above operations generate a three-dimensional energy tensor E with the same dimensions as the input tensor.

[0089] Sub-step 5.3 Nonlinear activation mapping and feature reweighting.

[0090] Finally, the computer inputs the calculated 3D energy tensor E into the Sigmoid nonlinear activation function, smoothly mapping its values ​​to the (0,1) interval to generate a clean attention weight mask. This weight mask is then compared with the original feature tensor to be cleaned. Element-wise multiplication (Hadamard Product) amplifies high-energy signal neurons while strongly suppressing low-energy background noise neurons, ultimately outputting a purified deep feature tensor. Its core mathematical formula is as follows: ; ; ; ; in, H represents the feature tensor to be cleaned output by the backbone network or neck network, which is a known quantity; H and W represent the height and width of the feature tensor, respectively, which are known quantities. These represent the spatial index coordinates of the tensor on the height and width axes, respectively, and are intermediate quantities for computation. These represent the mean and variance of spatial neuron activations of the feature tensor within a single channel, respectively, and are intermediate unknowns. This represents the minimum regularization coefficient preset by the system, and is a preset constant. This represents the three-dimensional energy tensor of the neuron, calculated using a closed-form mathematical solution, and is an intermediate unknown. This represents the Sigmoid nonlinear activation function operator, which is a preset constant. This represents the element-wise product operator based on the tensor broadcast mechanism, and is a preset constant; This represents the purified tensor, which effectively eliminates background false alarms and retains high-energy signal characteristics after three-dimensional energy reweighting, and is the quantity to be determined.

[0091] Step 6: Input the shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor into the decoupled detection head after multi-scale feature fusion, output pixel-level detection results, and output early warning messages based on the pixel-level detection results.

[0092] This step specifically includes the following sub-steps.

[0093] Sub-step 6.1: Multi-scale feature fusion and scale allocation.

[0094] The computer device will use the shallow high-resolution feature tensor obtained in step 204 The final complementary feature tensor output in step 204 and the purification feature tensor output in step 205 The inputs are fed into the Neck Network. This network performs scale-specific pruning, discarding computationally expensive deep-scale downsampling branches and outputting only a mesoscale feature tensor through upsampling and concatenation of the feature pyramid. With small-scale feature tensors .

[0095] Sub-step 6.2: Decoupling regression of the time-frequency boundary in pixel space.

[0096] The computer device combines the mesoscale feature tensor and the small-scale feature tensor. Input to the decoupled detection head operator. The decoupled detection head performs category prediction and spatial coordinate regression in parallel at each spatial grid point, outputting pixel-level detection results. These pixel-level detection results include pixel-level geometric regression bounding boxes. (Including pixel center coordinates) With pixel width and height ), Target existence confidence level and the probability distribution vector of drone model categories .

[0097] Sub-step 6.3: Physical reverse geometric mapping and business decision-making closed loop.

[0098] The computer equipment retrieves the short-time Fourier transform prior parameters set during the data preprocessing stage (this embodiment two), including the sampling rate. Time window step Number of points in Fourier transform The system utilizes inverse geometric mapping functions. ,Will Mapped to the real physical center time Duration of existence ;Will Mapped to physical center frequency With bandwidth usage Thus Transformed into a physical detection bounding box with practical electromagnetic significance. .

[0099] Finally, to eliminate overlapping boxes of the same target in adjacent grids, the system performs a check on all generated boxes. and its corresponding and Non-maximum suppression (NMS) algorithm is applied. The filtered single highest-confidence physical bounding box and category information are packaged into a standard early warning message. The data is then output to the upper-level operational command system to complete the detection loop. Its core algebraic continuation and mapping formulas are as follows: ; ; ; ; in, This represents the high-dimensional pure features extracted in the preceding steps, which are the known input quantities for the current step; This represents the shallow high-resolution feature tensor intercepted in the middle of the backbone network, which is a known input quantity; This represents the neck network operator that performs multi-scale pyramid fusion, and is a preset constant. , respectively, represent the mesoscale and small-scale feature tensors generated after dual-scale specific pruning and fusion, which are intermediate unknowns; This represents the decoupled detection head operator that performs classification and regression tasks, and is a preset constant. This represents the dimensionless pixel-level regression bounding box output at the end of the network (by...). (Construction), serving as intermediate quantities for computation; These represent the confidence level and model category probability distribution of the target predicted by the network, respectively, and are intermediate quantities in the calculation; These represent the inherent physical prior configurations of the system's sampling rate, step size, and FFT point count, respectively, and are preset constants; This represents the geometric mapping transformation function that reverses the analysis from pixel space to the real physical time-frequency space, and is a preset constant; This represents the detected bounding box that contains absolute physical meaning after decoding. Specifically, subscript Represents the physical center coordinates, prefix Represents the span. That is... For physical center time, The physical center frequency, For signal duration, To consume bandwidth; : Represents the non-maximum suppression algorithm operator used to remove spatially overlapping redundant boxes, which is a preset constant; This indicates that after redundancy removal and parameter packaging, the final end-to-end high-confidence warning command and decision message are output to the business system.

[0100] Example 2 To enable those skilled in the art to fully implement and reproduce this application, this embodiment provides an end-to-end engineered implementation plan, covering dataset construction, dedicated training strategies, performance verification data, and edge quantization deployment steps.

[0101] 7.1 Construction of radio frequency time-frequency datasets for complex electromagnetic environments.

[0102] 7.1.1 Radio frequency signal acquisition and scenario simulation.

[0103] Raw radio frequency signals were acquired in an open electromagnetic environment using a software-defined radio (SDR) receiver front-end. The dataset encompassed target signals from multi-rotor UAVs (such as remote control frequency hopping and digital image transmission), co-channel interference signals (such as Wi-Fi and Bluetooth), and ambient noise. Training samples with a signal-to-noise ratio ranging from 20 dB to -10 dB were constructed using digital mixing to enhance the model's ability to capture weak signals.

[0104] 7.1.2 Physical-level time-frequency conversion and dynamic range locking.

[0105] Perform a short-time Fourier transform (STFT) on the original baseband in-phase quadrature data sequence and hard truncate it using a fixed physical power threshold: ; Among them, the lower limit threshold of system noise The preferred range is -150dB to -130dB (preferably -140dB), signal energy upper limit threshold. The preferred range is -110dB to -90dB (preferably -100dB). This operation tightly locks the energy reference and effectively suppresses the abnormal amplification of pure background noise.

[0106] 7.2 Dedicated network training strategies for radio frequency signal topology characteristics.

[0107] 7.2.1 Joint loss function configuration.

[0108] End-to-end optimization is performed using a multi-task joint loss function: ; The weights of each loss term were experimentally calibrated, and the following optimal ranges were set: Classification loss weights The value range is 0.3-0.7 (preferably 0.5). Focal Loss is preferred, with its focusing parameters... The preferred range is 1.5-2.5; Target confidence loss weight The value range is 0.5-0.9 (preferably 0.7). Binary cross-entropy loss is preferred. Bounding box regression loss weights The value range is 0.03-0.08 (preferably 0.05). CIoULoss is preferred to accurately fit the extremely high aspect ratio topology unique to radio frequency signals.

[0109] 7.2.2 Training hyperparameter settings and radio frequency enhancement strategies.

[0110] In a preferred embodiment, the network training configuration is as follows: Batch Size: Values ​​range from 8 to 32 (preferably 16); Initial learning rate: ranges from 0.0005 to 0.005 (preferably 0.001), and is dynamically scheduled in conjunction with a cosine annealing strategy; Optimizer: AdamW is preferred, with its weight decay coefficient set between 0.01 and 0.05 (preferably 0.025). Training epochs: 80-150 epochs (preferably 120 epochs).

[0111] Meanwhile, radio frequency-specific data enhancements are introduced during training: including random erasure in the time-frequency domain (randomly selecting 10%-30% of the region in the time-frequency spectrum to be zeroed, simulating the signal being blocked by instantaneous strong interference), band-limited noise injection (simulating narrowband interference in a specific frequency band), and mixed enhancement of multiple signal-to-noise ratios in the same batch.

[0112] 7.3 Example of offline model performance verification.

[0113] To objectively evaluate the effectiveness of this application, performance evaluation and comparative verification were conducted on an offline training platform (such as a server equipped with a high-performance GPU).

[0114] 7.3.1 Benchmark model comparison and verification.

[0115] The performance comparison results of this application's model and existing mainstream lightweight detection networks on the same radio frequency dataset are as follows: High signal-to-noise ratio (Clean) environment performance: The model in this application has only 0.83M parameters and a multiply-accumulate computational cost (GFLOPs) of only 5.5. Compared to the benchmark model YOLOv5n (2.50M parameters, 7.1 GFLOPs), this application achieves an mAP50 of 0.9924 while reducing the number of parameters by approximately 66.8%. In a GPU environment, real-time inference speed can reach over 249 FPS.

[0116] Robustness in low signal-to-noise ratio environments: Under strong noise background of 0dB, the mAP50 of this application reaches 0.4201, which is significantly better than YOLOv5n (0.3627), YOLOv8n (0.2680) and YOLO11n (0.1863), fully verifying the enhanced extraction capability of this application for weak radio frequency targets.

[0117] 7.3.2 Ablation experiment verification of core modules.

[0118] To verify the effectiveness of the anisotropic module design, ablation tests were conducted in a 0dB environment. Experiments show that compared to the conventional baseline network (mAP50: 0.3627), the detection accuracy improved to 0.3854 after introducing the AP-Pzconv operator; further, the accuracy jumped to 0.4012 after adding the TF-CCG mechanism; finally, after incorporating the parameterless energy attention cleansing mechanism, the system's mAP50 reached its optimal value of 0.4201.

[0119] 7.4 Quantitative Deployment Examples for Edge Computing Devices

[0120] To deploy the above model to portable detection devices with limited computing power and power consumption (such as processing systems with built-in NPUs or ARM architectures), the computer equipment further performs the following deployment optimization steps: 1. Computation graph optimization and operator fusion: Remove redundant nodes in the training phase and perform mathematical folding (Conv-BN Fusion) on the deep convolutional layer (DCV) and batch normalization layer (BN) in the feature extraction network to reduce memory access overhead during inference. 2. Model compression and quantization conversion: Using the calibration dataset, the network weights are quantized from floating-point precision (FP32) to integer precision (INT8), which significantly compresses the model size while maintaining the aforementioned detection accuracy.

[0121] 3. Heterogeneous pipelined inference design: On edge devices, the STFT time-frequency conversion is assigned to the CPU or DSP for execution, and the forward propagation of the quantized neural network is assigned to a dedicated hardware acceleration unit (such as NPU / GPU) for execution, thereby realizing an end-to-end real-time RF detection service closed loop under low power consumption.

[0122] Example 3 This application has the following technical advantages: I. Detection efficiency and computing resource optimization: 1. Technical Effects of the Anisotropic Feature Extraction Structure Based on AP-Pzconv: Addressing the directional sensitivity of UAV RF signals in the time-spectrum image, this application employs an anisotropic feature extraction structure combining channel decoupling and asymmetric directional convolution to establish directional receptive fields along both the time and frequency axes. Specifically, lateral directional convolution is used to model the temporal continuity features of frequency-hopping signals, while longitudinal directional convolution is used to model the frequency spread features of broadband signals, thereby reducing redundant computation in invalid background regions by traditional isotropic convolution. Furthermore, through partial channel convolution and depthwise separable convolution structures, the number of model parameters is reduced to 0.83M, a reduction of approximately 66.8% compared to YOLOv5n's 2.50M parameters; GFLOPs are reduced from 7.1 to 5.5, improving edge device deployment capabilities while maintaining detection performance.

[0123] Experimental results show that, under the Clean dataset conditions, the mAP50 of the model in this application reaches 0.9924, and the mAP50-95 reaches 0.9646; even under a low signal-to-noise ratio environment of 0dB, the mAP50 still reaches 0.4201, verifying the stable feature extraction capability of the structure in complex electromagnetic environments.

[0124] 2. Technical Effects of the Multi-Scale Feature Pruning Strategy: Based on the scale distribution characteristics of UAV radio frequency targets in the time-spectrum image, this application prunes the high-downsampled deep feature branches while retaining the mid-to-low-level feature fusion structure. Furthermore, computational resources are concentrated on the parsing of small-to-medium scale target features, which reduces the computational complexity of the model while improving the real-time inference capability in edge device scenarios.

[0125] Experimental results show that the model in this application achieves an inference speed of over 249 FPS in edge device scenarios, meeting the real-time detection requirements while maintaining high detection accuracy.

[0126] II. Enhanced target continuity in complex electromagnetic environments: 1. Technical Effects of the TF-CCG Time-Frequency Complementary Feature Reconstruction Mechanism: Addressing the issue of fragmented and discontinuous radio frequency features in UAVs operating under complex electromagnetic environments, this application employs a bidirectional cross-constraint mechanism between semantic and positional feature branches to generate complementary guiding masks. Positional features constrain the deep semantic response region, while semantic features enhance the continuity of the shallow positional region, thereby achieving continuous reconstruction of fragmented time-frequency features.

[0127] Experimental results show that, in a 0dB low signal-to-noise ratio environment, the mAP50 of the model in this application reaches 0.4201, which is about 15.8% higher than YOLOv5n's 0.3627; the mAP50-95 reaches 0.3806, which is significantly higher than YOLOv5n's 0.2565, verifying the enhancement capability of this mechanism for weak radio frequency targets in complex noise environments.

[0128] 2. Technical Effects of the Striped Directional Context Aggregation Mechanism: This application performs directional context aggregation operations along both the time and frequency axes to extract long-range dependencies of radio frequency signals in the time and frequency dimensions. Compared to global average pooling, this method can preserve the directional topological information in the UAV's radio frequency time-frequency spectrum and enhance the contextual correlation ability between broken signal regions. Furthermore, this mechanism can enhance the continuous representation ability of time-frequency structures in complex background environments and improve the target identification stability under low signal-to-noise ratio conditions.

[0129] III. Enhanced model generalization ability in complex background environments: 1. Technical advantages of dynamic range mapping based on physical threshold constraints: To address the issue that traditional dynamic range normalization can easily amplify background thermal noise, this application introduces a fixed physical power threshold to truncate the time-spectrum graph. Background noise regions below the lower power threshold are suppressed, while signal regions above the threshold are preserved, thereby reducing the impact of noise disturbances on the network input in complex background environments. Furthermore, this improves the consistency of the background distribution of the time-spectrum graph under different electromagnetic environmental conditions, enhancing the model's generalization ability in unknown and complex noise environments.

[0130] 2. Technical effects of the parameter-free energy noise suppression mechanism: This application preferably employs a parameter-free energy attention mechanism to suppress background noise responses in deep feature maps. By independently evaluating the energy of neurons in the feature space, it enhances high-response target regions and reduces the interference from low-energy background regions.

[0131] Experimental results show that, without adding any additional learnable parameters, the introduction of this mechanism improves the model's mAP50 from 0.1197 to 0.4201 in a 0dB low signal-to-noise ratio environment, verifying the mechanism's ability to enhance weak radio frequency targets in complex background environments.

[0132] IV. Experimental Verification: To verify the effectiveness of the method in this application, the model in this application was compared with lightweight target detection models such as YOLOv5n, YOLOv8n, and YOLO11n on the same UAV radio frequency dataset.

[0133] Experimental results show that, under the Clean dataset conditions, the model in this application achieves an mAP50 of 0.9924 and an mAP50-95 of 0.9646, while maintaining high detection accuracy and having only 0.83M model parameters.

[0134] Furthermore, in a low signal-to-noise ratio (SNR) environment of 0 dB, the mAP50 of the model in this application reaches 0.4201, which is higher than 0.3627 for YOLOv5n, 0.2680 for YOLOv8n, and 0.1863 for YOLO11n, verifying the robust detection capability of this application for weak UAV radio frequency targets in complex electromagnetic environments.

[0135] Meanwhile, the model in this application reduces GFLOPs to 5.5 and achieves an inference speed of over 249 FPS, which can meet the real-time deployment requirements in edge device scenarios.

[0136] Example 4 In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores and processes data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a UAV radio frequency detection method based on anisotropy and time-frequency complementarity.

[0137] Figure 6 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0138] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0139] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0140] It should be noted that 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 related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0141] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0142] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0144] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for radio frequency detection of unmanned aerial vehicles (UAVs) based on anisotropy and time-frequency complementarity, characterized in that, The method includes: Acquire spatial analog radio frequency signals, and perform down-conversion and analog-to-digital conversion on the spatial analog radio frequency signals to obtain discrete baseband in-phase quadrature data sequences; A short-time Fourier transform is performed on the baseband in-phase orthogonal data sequence to generate a two-dimensional logarithmic power spectrum. Based on the preset physical absolute power lower threshold and physical absolute power upper threshold, the two-dimensional logarithmic power spectrum is dynamically range hard-truncated to generate a single-channel normalized tensor. The single-channel normalized tensor is then processed to obtain a three-channel anisotropic time-frequency tensor. First, the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network is used as the first input feature tensor. The first input feature tensor is then decoupled along the channel dimension in an asymmetric proportion into a time feature flow tensor, a frequency feature flow tensor, and an identity-preserving flow tensor. Next, one-dimensional convolutions with different directions are performed on the time feature flow tensor and the frequency feature flow tensor to obtain the one-dimensional convolution results. Finally, the one-dimensional convolution results are concatenated with the identity-preserving flow tensor along the channel dimension, and the concatenated tensor is subjected to pointwise convolution and residual connection to output the final anisotropic combined feature tensor. First, the final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted from the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor. The local feature tensor is then decoupled into a semantic initial flow tensor and a position initial flow tensor. Next, specific mapping is performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate a high-dimensional semantic tensor and a high-dimensional position tensor. Then, a dual-path orthogonal global pooling operation is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate a semantic attention mask tensor and a position attention mask tensor. Finally, the semantic attention mask tensor is multiplied element-wise with the high-dimensional position tensor to obtain the first multiplication result. The position attention mask tensor is multiplied element-wise with the high-dimensional semantic tensor to obtain the second multiplication result. Finally, the first multiplication result and the second multiplication result are added together and fused by pointwise convolution to output the final complementary feature tensor of break-reconnection. Based on the final complementary feature tensor, the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space is calculated. After nonlinear activation, energy attention weights are generated, and the energy attention weights are multiplied element by element with the final complementary feature tensor to output the purified feature tensor after background noise suppression. The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are fused at multiple scales and then input into the decoupled detection head to output pixel-level detection results. Based on the pixel-level detection results, early warning messages are output.

2. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, Based on preset physical absolute power lower and upper threshold values, a dynamic range hard truncation is performed on the two-dimensional logarithmic power spectrum to generate a single-channel normalized tensor, specifically: ； in, This represents the single-channel normalized tensor generated after hard truncation; Represents a two-dimensional logarithmic power spectrum; This represents the lower limit threshold of physical absolute power. This represents the upper limit threshold of physical absolute power.

3. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, The one-dimensional convolution result includes a horizontally mapped tensor and a vertically mapped tensor. First, the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network is used as the first input feature tensor. This first input feature tensor is then decoupled along the channel dimension in an asymmetric proportion into a time feature component tensor, a frequency feature component tensor, and an identity-preserving component tensor. Next, one-dimensional convolutions with different directions are performed on the time and frequency feature component tensors to obtain the one-dimensional convolution result. Finally, the one-dimensional convolution result is concatenated with the identity-preserving component tensor along the channel dimension, and the concatenated tensor is subjected to pointwise convolution and residual concatenation to output the final anisotropic combined feature tensor, specifically including: Use the three-channel anisotropic time-frequency tensor or the intermediate layer feature tensor in the backbone network as the first input feature tensor; The first input feature tensor is decoupled according to the channel ratio of 1:1:

2. The first quarter of the channels are extracted as the time feature flow tensor, the middle quarter of the channels are extracted as the frequency feature flow tensor, and the remaining half of the channels are used as the identity-preserving flow tensor. Perform a one-dimensional horizontal spatial convolution with a kernel of K×1 on the temporal feature flow tensor to generate a horizontal mapping tensor, and perform a one-dimensional vertical spatial convolution with a kernel of K×1 on the frequency feature flow tensor to generate a vertical mapping tensor; where K is an odd number of 5, 7 or 9. The horizontal mapping tensor, the vertical mapping tensor, and the identity-preserving flow tensor are concatenated along the channel dimension to obtain the concatenated tensor. Then, pointwise convolution and residual connection are performed on the concatenated tensor to output the final anisotropic combined feature tensor.

4. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, The final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted from the midstream of the backbone network are concatenated through channel splicing to form a local feature tensor. This local feature tensor is then decoupled into a semantic initial flow tensor and a positional initial flow tensor, specifically including: The final anisotropic combined feature tensor and the shallow high-resolution feature tensor intercepted in the middle of the backbone network are concatenated through channel splicing to form a local feature tensor; The local feature tensor is decoupled along the channel dimension at a ratio of 1:

3. The first quarter of the channels is extracted as the semantic initial flow tensor, and the remaining three-quarters of the channels are used as the position initial flow tensor.

5. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, Specific mappings are performed on the semantic initial flow tensor and the position initial flow tensor respectively to generate high-dimensional semantic tensors and high-dimensional position tensors. Then, bi-path orthogonal global pooling operations are performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate semantic attention mask tensors and position attention mask tensors. Specifically, this includes: The semantic initial flow tensor is subjected to two consecutive spatial convolutions followed by channel expansion convolutions to generate a high-dimensional semantic tensor. In parallel, the position initial flow tensor is subjected to pointwise convolutions to generate a high-dimensional position tensor. Deep convolution is performed on the high-dimensional semantic tensor and the high-dimensional position tensor respectively to generate semantic local conditioning tensor and position local conditioning tensor; The semantic local conditioning tensor and the positional local conditioning tensor are subjected to dual-path spatial adaptive average pooling and element-wise multiplication, respectively, to obtain the semantic global interaction tensor and the positional global interaction tensor. Pointwise convolution and nonlinear activation function mapping are performed on the semantic global interaction tensor and the positional global interaction tensor respectively to generate the semantic attention mask tensor and the positional attention mask tensor.

6. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, Based on the final complementary feature tensor, the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space is calculated, specifically: ； in, This represents the ratio of the square of the deviation of each neuron from the mean of its channel space to the variance of its channel space. Represents the final complementary feature tensor; This indicates that each neuron deviates from the mean of its channel space. ; This represents the variance of each neuron's deviation from the spatial distribution of its channel. ; H represents the preset minimum regularization coefficient; H and W represent the height and width of the final complementary feature tensor, respectively. This indicates retrieving all channels; These represent the spatial index coordinates of the final complementary feature tensor on the height and width axes, respectively.

7. The UAV radio frequency detection method based on anisotropy and time-frequency complementarity according to claim 1, characterized in that, The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are fused at multiple scales and then input into the decoupled detection head. Pixel-level detection results are output, and based on these results, early warning messages are generated, including: The shallow high-resolution feature tensor, the final complementary feature tensor, and the cleaned feature tensor are input into the neck feature fusion network, which outputs a mesoscale feature tensor and a small-scale feature tensor. The neck feature fusion network is used for upsampling and splicing fusion through the feature pyramid. The mesoscale and small-scale feature tensors are input into the decoupled detection head, and class prediction and spatial coordinate regression are performed in parallel at each spatial grid point, outputting pixel-level detection results. The pixel-level detection results include pixel-level geometric regression bounding boxes, target presence confidence, and UAV model category probability distribution vectors. Based on preset short-time Fourier transform prior parameters, the pixel-level geometric regression bounding box is processed using the inverse geometric mapping function to generate a physical detection bounding box. The nonmaximum suppression algorithm is used to process the physical detection bounding box and the corresponding target existence confidence and UAV model category probability distribution vector to obtain the early warning message.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the UAV radio frequency detection method based on anisotropy and time-frequency complementarity as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the UAV radio frequency detection method based on anisotropy and time-frequency complementarity as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the UAV radio frequency detection method based on anisotropy and time-frequency complementarity as described in any one of claims 1-7.

Images