Method, device and equipment for constructing electromagnetic spectrum for unmanned aerial vehicle swarm detection

Abstract

This application relates to a method, apparatus, and device for constructing electromagnetic maps for UAV swarm detection. The method includes: acquiring mixed baseband signals from a distributed multi-observation station simultaneously collected by the UAV swarm; extracting a one-dimensional time-domain observation sequence for each array element; and generating a time-frequency image through time-frequency transformation. The image is input into a time-frequency signal parameter extraction network to extract the time-domain start-end information, center frequency, and physical bandwidth of each signal component. Based on the time-domain start-end information, all observation matrices are simultaneously subjected to time-domain hard truncation slices, decoupling the mixed baseband signals into quasi-single-target signal segments. Three-dimensional localization is performed on each segment to obtain preliminary coordinates, and unsupervised spatial clustering is performed based on coordinate distance to divide radiation source clusters. Finally, the three-dimensional coordinates and electromagnetic parameters of signals from the same radiation source are fused to generate a complete electromagnetic map. This method enables blind separation of UAV swarm signals, multi-target association without prior knowledge, and autonomous construction of high-precision multi-dimensional electromagnetic maps.

Description

Technical Field

[0001] This application relates to the field of signal processing technology, and in particular to a method, apparatus and device for constructing electromagnetic spectrum for UAV swarm detection. Background Technology

[0002] Drone swarms typically employ a distributed, collaborative working mode, exhibiting high mobility and strong stealth. In complex, non-cooperative electromagnetic environments, the radiation signals from swarms often display complex characteristics such as multi-source aliasing, high bursts, wide bandwidth, and dynamic variations in time and frequency parameters, posing a severe challenge to existing electromagnetic space detection and situational awareness technologies.

[0003] Currently, traditional electromagnetic detection and situational awareness technologies have many limitations when dealing with the aforementioned complex scenarios. First, in terms of spectrum sensing and signal separation, traditional scanning receivers or analysis methods based on conventional energy detection are difficult to capture short-lived burst signals in real time over a wide frequency range. Furthermore, when multiple UAV radiation sources emit simultaneously, causing severe aliasing in the time and frequency domains, existing technologies lack efficient blind separation methods, making it difficult to accurately cut out and extract the time and frequency characteristic parameters of a single radiation source from the mixed baseband data.

[0004] In target spatial localization, traditional two-step localization methods based on angle of arrival or time difference of arrival inevitably suffer from problems such as difficulty in matching parameters at multiple stations and spurious correlations when dealing with multiple targets. Although direct localization technology can alleviate the parameter matching problem, it is difficult to directly map aliased signals into accurate three-dimensional spatial coordinates under blind detection conditions where prior information is missing and signals overlap, resulting in high computational difficulty and a high likelihood of localization confusion.

[0005] At the system-level fusion and map construction level, existing electromagnetic detection equipment typically treats frequency domain parameter measurement and spatial direction finding in isolation, lacking an effective cross-dimensional spatiotemporal fusion mechanism. When faced with massive, discrete short-time signal fragments, existing technologies lack unsupervised feature correlation and clustering capabilities, and cannot automatically match scattered electromagnetic parameters and spatial coordinates to corresponding physical entities when the number of targets is unknown.

[0006] Therefore, how to solve the detection and positioning problems caused by the aliasing of signals from multiple radiation sources in complex electromagnetic environments without prior information, and how to achieve high-precision multi-dimensional electromagnetic situational awareness in UAV swarm detection scenarios, has become an urgent technical problem to be solved. Summary of the Invention

[0007] Therefore, it is necessary to provide a method, apparatus, and device for constructing electromagnetic maps for UAV swarm detection, addressing the aforementioned technical problems.

[0008] A method for constructing electromagnetic maps for UAV swarm detection, the method comprising: Acquire the hybrid baseband signal of the target UAV swarm, which is synchronously collected by multiple distributed observation stations; the hybrid baseband signal includes the multi-element observation matrix of each observation station; Extract a one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image; The time-frequency image is input into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; Based on the time-domain start and end information, the multi-element observation matrix of each observation station is synchronously hard-truncated in the time domain and sliced ​​to decouple the mixed baseband signal into multiple quasi-single target signal segments. The quasi-single target signal segment is spatially located in three dimensions to obtain the corresponding preliminary three-dimensional coordinates. The spatial distance between all preliminary three-dimensional coordinates is calculated, and unsupervised spatial clustering is performed based on the spatial distance to obtain multiple clusters. Each cluster corresponds to a radiation source. The three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source are statistically fused to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and the physical bandwidth.

[0009] An electromagnetic map construction device for UAV swarm detection, the device comprising: The signal acquisition module is used to acquire the hybrid baseband signal of the target UAV swarm simultaneously acquired by distributed multi-observation stations; the hybrid baseband signal includes the multi-element observation matrix of each observation station; The time-frequency transformation module is used to extract a one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image. The parameter extraction module is used to input the time-frequency image into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; The signal slicing module is used to perform synchronous time-domain hard truncation slicing on the multi-element observation matrix of each observation station according to the time-domain start and end information, and decouple the mixed baseband signal into multiple quasi-single target signal segments. The clustering module is used to perform three-dimensional spatial localization on the quasi-single target signal segment, obtain the corresponding preliminary three-dimensional coordinates, calculate the spatial distance between all preliminary three-dimensional coordinates, and perform unsupervised spatial clustering based on the spatial distance to obtain multiple clusters; each cluster corresponds to a radiation source. The spectrum generation module is used to statistically fuse the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and the physical bandwidth.

[0010] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: Acquire the hybrid baseband signal of the target UAV swarm, which is synchronously collected by multiple distributed observation stations; the hybrid baseband signal includes the multi-element observation matrix of each observation station; Extract a one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image; The time-frequency image is input into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; Based on the time-domain start and end information, the multi-element observation matrix of each observation station is synchronously hard-truncated in the time domain and sliced ​​to decouple the mixed baseband signal into multiple quasi-single target signal segments. The quasi-single target signal segment is spatially located in three dimensions to obtain the corresponding preliminary three-dimensional coordinates. The spatial distance between all preliminary three-dimensional coordinates is calculated, and unsupervised spatial clustering is performed based on the spatial distance to obtain multiple clusters. Each cluster corresponds to a radiation source. The three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source are statistically fused to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and the physical bandwidth.

[0011] The aforementioned electromagnetic spectrum construction method, apparatus, and equipment for UAV swarm detection, by extracting the time-domain sequence of a single array element and converting it into a time-frequency image, can transform one-dimensional aliased signals into a two-dimensional visualization form capable of feature mining. This is beneficial for highlighting the characteristics of weak UAV signals under low signal-to-noise ratio conditions. Based on the extracted time-domain start and end information, global synchronous time-domain hard truncation slicing can decouple multi-source overlapping and aliased signals into independent quasi-single-target segments, effectively reducing signal coupling interference. By combining quasi-single-target segment three-dimensional positioning with unsupervised spatial clustering, it is possible to complete the entity attribution of discrete positioning points without pre-setting the number of radiation sources, achieving blind estimation of the number of radiation sources. By statistically fusing multi-dimensional parameters of the same radiation source to generate an electromagnetic spectrum, it can suppress random errors in single detections and form a complete situational characterization that combines spatial location and electromagnetic attributes. In this embodiment of the invention, blind separation of UAV swarm signals, multi-target association without prior knowledge, and autonomous construction of high-precision multi-dimensional electromagnetic spectra can be achieved in complex electromagnetic environments with low signal-to-noise ratio and multi-source aliasing. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating an electromagnetic spectrum construction method for UAV swarm detection in one embodiment. Figure 2 This is a schematic diagram of the system output situation at t=0s in one embodiment; Figure 3 This is a schematic diagram of the system output situation at t=20s in one embodiment; Figure 4 This is a schematic diagram of the system output situation at t=40s in one embodiment; Figure 5 This is a schematic diagram of the system output situation at t=60s in one embodiment; Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] In one embodiment, such as Figure 1 As shown, a method for constructing electromagnetic maps for UAV swarm detection is provided, including the following steps: Step 102: Acquire the hybrid baseband signal from the target UAV swarm, which is synchronously collected by distributed multi-observation stations. The hybrid baseband signal includes the multi-element observation matrix of each observation station.

[0015] Distributed multi-observation stations are multiple passive observation base stations that are distributed in space and have high-precision time synchronization capabilities. The hybrid baseband signal is the aliased baseband data formed by the simultaneous reception of multiple radiation sources from the UAV swarm by the antenna array elements of each observation base station. The multi-element observation matrix is ​​a two-dimensional array data matrix composed of all antenna array elements of a single observation base station at continuous sampling times.

[0016] Step 104: Extract the one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image.

[0017] The one-dimensional time-domain observation matrix is ​​a single-channel time-domain sampling sequence separated from a single array element in the multi-element observation matrix. The time-frequency image is a gray-scale feature image that can intuitively reflect the time, frequency, and energy distribution of the signal after normalization processing.

[0018] Step 106: Input the time-frequency image into the pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component.

[0019] The time-frequency signal parameter extraction network is a lightweight deep learning network trained offline and adapted for time-frequency image feature analysis, and a lightweight YOLO series network can be used. The signal components are independent burst signal regions of UAV radiation sources in the time-frequency image. The time-domain start and end information is the sampling start and end position corresponding to a single signal component. The center frequency and physical bandwidth are the carrier center frequency and effective occupied frequency band range corresponding to the signal component, respectively.

[0020] It is understandable that by automatically parsing time-frequency images and outputting multiple types of electromagnetic feature parameters through a time-frequency signal parameter extraction network, blind detection and accurate parameter inversion of multiple signal components in aliased scenarios can be achieved without prior human intervention, which is beneficial to improving the real-time performance and automation of signal feature extraction in complex electromagnetic environments.

[0021] Step 108: Based on the start and end information in the time domain, perform synchronous time domain hard truncation slicing on the multi-element observation matrix of each observation station to decouple the mixed baseband signal into multiple quasi-single target signal segments.

[0022] Synchronous time-domain hard truncation slicing refers to a processing method that uses unified time-domain start and end information as the cutting benchmark to synchronously truncate and extract the observation matrices of all observation stations in the time dimension. Quasi-single target signal segments are array data segments that have been significantly weakened by interference after time-domain slicing and approximately contain the radiation source signal of a single UAV.

[0023] It is understandable that by implementing global synchronous time-domain slicing based on the extracted time-domain start and end information, it is possible to decouple and reduce the dimensionality of composite signals with severe multi-source overlap, and separate them into mutually independent quasi-single target signal segments. This is beneficial to avoid coupling interference between multiple signals and reduce the difficulty of subsequent three-dimensional spatial positioning calculations.

[0024] Step 110: Perform three-dimensional spatial localization on the single target signal segment to obtain the corresponding preliminary three-dimensional coordinates. Calculate the spatial distance between all preliminary three-dimensional coordinates and perform unsupervised spatial clustering based on the spatial distance to obtain multiple clusters. Each cluster corresponds to a radiation source.

[0025] Three-dimensional spatial positioning is an array signal processing method that uses array observation data to invert and solve the three-dimensional geographic coordinates of radiation sources. The initial three-dimensional coordinates are discrete spatial location points obtained by solving a single quasi-single target signal segment. Unsupervised spatial clustering is a processing method that automatically aggregates and divides discrete positioning points based on spatial distance similarity under the premise of unknown number of radiation sources. The cluster is a set of positioning points that meet the spatial distance association threshold and belong to the same UAV radiation source.

[0026] It is understandable that by combining segment-by-segment 3D localization with spatial distance-driven unsupervised clustering, it is possible to automatically classify the entity affiliation of discrete localization points and blindly estimate the number of radiation sources without pre-setting the number of targets, which is beneficial to improving the reliability of multi-target identity matching and individual differentiation of UAV swarms.

[0027] Step 112 involves statistically fusing the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum. The electromagnetic parameters include the center frequency and physical bandwidth.

[0028] Statistical fusion is a method of integrating multiple sets of discrete three-dimensional coordinates and electromagnetic parameters corresponding to the same radiation source by smoothing and averaging. Electromagnetic spectrum is an electromagnetic spatial situation representation spectrum formed by integrating the spatial location and electromagnetic properties of the radiation source.

[0029] It is understandable that by statistically fusing the spatial coordinates and electromagnetic parameters of signal segments within the same cluster, the random errors of single positioning and parameter detection can be reduced, forming a stable and reliable multidimensional feature fingerprint of radiation sources. This is beneficial for constructing a complete electromagnetic situation map that combines spatial location and frequency domain electromagnetic properties, supporting continuous situational awareness of UAV swarms across the entire domain.

[0030] In the aforementioned electromagnetic map construction method for UAV swarm detection, by extracting the time-domain sequence of a single array element and converting it into a time-frequency image, one-dimensional aliased signals can be transformed into a two-dimensional visualization form that can be used for feature mining. This is beneficial for highlighting the characteristics of weak UAV signals under low signal-to-noise ratio conditions. Based on the extracted time-domain start and end information, global synchronous time-domain hard truncation slicing can be implemented, decoupling multi-source overlapping and aliased signals into independent quasi-single-target segments, effectively reducing signal coupling interference. By combining 3D localization of quasi-single-target segments with unsupervised spatial clustering, the entity attribution of discrete positioning points can be completed without pre-setting the number of radiation sources, achieving blind estimation of the number of radiation sources. By statistically fusing multi-dimensional parameters of the same radiation source to generate an electromagnetic map, random errors in single detection can be suppressed, forming a complete situational characterization that combines spatial location and electromagnetic properties. This invention enables blind separation of UAV swarm signals, multi-target association without prior knowledge, and autonomous construction of high-precision multi-dimensional electromagnetic maps in complex electromagnetic environments with low signal-to-noise ratios and multi-source aliasing.

[0031] In one embodiment, the method of the present invention includes the deployment of a distributed multi-station array passive detection hardware architecture.

[0032] Specifically, distributed deployment in a wide-area detection space The first fixed observation base station in the system that is strictly separated in space. The three-dimensional coordinates of each base station are defined as follows: Each base station is equipped with a uniform circular array consisting of L omnidirectional antenna elements, the physical radius of which is R. For Q far-field UAV radiation source targets in space, arranged in formation and radiating unknown broadband radio frequency signals, their true physical three-dimensional coordinates are: Each base station's discrete-time index The receiving channel is simultaneously activated. Considering the combined effects of free-space channel attenuation, transmission phase delay, and additive environmental thermal noise, the first element containing the output characteristics of all array elements is intercepted and generated. Multi-source aliasing global complex baseband observation signal matrix of each base station This provides the basic input data for the subsequent separation of blind signals.

[0033] The receiving node generates the path fading response and complex baseband signal matrix for a free-space wavefront: According to classical electromagnetic theory, the logarithmic mathematical model of electromagnetic attenuation during free-space physical path transmission must be rigorously quantified as follows: ; in, The free-space path loss between the m-th observation station and the radiation source is... Let m be the three-dimensional linear propagation distance from the m-th observation base station to the radiation source. The carrier frequency of the signal emitted by the radiation source. It is the speed of light in a vacuum.

[0034] Considering the thermodynamic power level constant of additive white Gaussian noise in the ambient environment The derivation formula for the actual intrinsic signal-to-noise ratio variable of the observation node relative to the radiation source channel link is as follows: ; in, Let m be the intrinsic signal-to-noise ratio of the receiving link at the m-th observation base station. The received signal power is the signal power from the radiation source that reaches the aperture of the receiving antenna. For the gain of the receiving antenna, For the gain of the transmitting antenna, The propagation distance from the radiation source to the observation node. The establishment of the fading model for the carrier frequency of the radiated signal, a spatial nonlinear physical quantity, not only ensures that the following baseband discrete sequence formula strictly follows the basic conservation laws of microwave transmission in nature, but also provides a mapping basis for the communication physical environment at the underlying logic level for channel fading simulation and algorithm verification during subsequent near-field / far-field switching processes. ; in, Let l be the complex baseband received signal sample value of the l-th antenna element of the m-th base station at the k-th discrete sampling time. Let be the received signal power corresponding to the l-th array element of the m-th base station. Let be the sampled value of the baseband signal emitted by the radiation source at the k-th sampling time. The phase factor is in complex exponential form. It represents the additive white Gaussian noise sample value of the l-th array element of the m-th base station at the k-th sampling time.

[0035] In one embodiment, performing time-frequency transformation on a one-dimensional time-domain observation sequence and generating a time-frequency image includes: performing a short-time Fourier transform on the one-dimensional time-domain observation sequence to obtain a two-dimensional time-frequency energy distribution; and performing logarithmic compression and linear scaling on the two-dimensional time-frequency energy distribution to generate a standardized grayscale time-frequency image.

[0036] Specifically, extracting a subset sequence of observation data from a reference node array element in the observation system. Using sliding window functions Applying a discrete short-time Fourier transform projects a one-dimensional pure time-domain sequence onto a two-dimensional time-frequency energy surface. In the middle. Take the square of its complex magnitude and perform a logarithmic compression transformation. A two-dimensional grayscale time-frequency image is generated. Simultaneously, the detection space is defined with absolute physical start and end times. The broadband signal component with absolute physical bandwidth B is precisely transformed into a normalized computer vision coordinate system through a linear compression ratio mathematical mapping, generating a set of normalized center coordinates and width and height parameters that are strongly decoupled from the image scale. ,in, The x-coordinate of the center of the normalized bounding box. The center ordinate of the normalized bounding box. To normalize the width of the bounding box, To normalize the height of the bounding box, the complex electromagnetic parameter estimation is transformed into a standard target bounding box visual regression task.

[0037] The core derivation formula of the discrete short-time Fourier transform based on sliding overlap processing is as follows: ; in, It is a two-dimensional time-frequency complex energy matrix. For indexing time windows, For frequency index, This represents the total number of sampling points in a single short-time Fourier transform (FFT points). This represents the complex baseband value at the nth sampling point. For sliding window functions, This is the index of the time-domain sampling point. The window sliding step size (frame shift) The imaginary unit is the exponential basis vector. Characterized by the projection weight integral of orthogonal frequency decomposition; smoothing window function variable The time of the Gaussian filter after effective spectrum cutoff is determined by a constant step scalar. This directly and strongly dominates the temporal window width resolution of the transformed energy image in the lateral extension direction. The inverse regression transformation mapping formula for its parameters is: ; in, This is the actual physical start time of the signal. The total duration of the entire observation window. The actual physical duration (time width) of the signal, and the formula: ; in, The actual physical center frequency of the signal. The maximum absolute frequency of the system analysis perfectly extends the physical Hertzian dimension and the second constant dimension into the pure 0 to 1 neural network perception real number domain without loss.

[0038] In one embodiment, the time-frequency signal parameter extraction network includes an edge-enhanced multi-scale feature extraction module and a frequency-space interaction detection head. The edge-enhanced multi-scale feature extraction module performs multi-scale feature extraction and high- and low-frequency component separation on the input time-frequency image, and outputs an enhanced multi-scale spatial feature map. The frequency-space interaction detection head performs orthogonal wavelet frequency domain lossless dimensionality reduction on the enhanced multi-scale spatial feature map to generate a frequency domain feature map. The spatial feature map and the frequency domain feature map are then fused and fused with dual-domain features and adaptive weight calibration. Finally, the normalized bounding box coordinates of each signal component are output through bounding box regression.

[0039] Specifically, the normalized time-frequency image is input into a pre-trained, offline-trained improved YOLOv8 deep learning detection network (i.e., the YOLOv8-LFSI-DH-EOMFE framework). This improved YOLOv8 deep learning detection network is an improvement on the native YOLOv8 architecture: in the feature extraction and fusion stage, the Edge-Optimized Multi-scale Feature Extraction (EOMFE) module replaces the original C2f structure; in the prediction stage, the Lightweight Frequency-Spatial Interaction Detection Head (LFSI-DH) replaces the original decoupled detection head. This network extracts noise-resistant spatial contours through the EOMFE module and performs frequency-adaptive lossless dimensionality reduction using orthogonal decomposition basis through LFSI-DH. The network performs end-to-end parallel forward inference. The network performs end-to-end parallel forward inference, outputting the predicted bounding box coordinates of each independent signal burst segment in the aliased scene, and performs physical scale inverse mapping parsing to accurately extract the temporal starting sampling point index corresponding to the k-th valid signal component event. index of time-domain termination sampling points Center frequency and the absolute bandwidth of the physical signal .

[0040] The internal algorithm mechanism of the EOMFE module for enhancing and extracting weak background edge features is limited to: the edge enhancer performs controlled suppression of target noise. The resulting multi-scale degradation matrix is ​​forced to undergo downscaling—the formula for joint 2D shared convolution is: ; in, This is the output feature map after two-stage convolution processing. This is the weight matrix of the first-level shared convolutional layer. This is the weight matrix of the second-level shared convolutional layer. The input feature map for the j-th layer is the multi-scale degraded feature matrix affected by noise. This is the bias vector for the first-level shared convolutional layer. The bias vector of the second-level shared convolutional layer is reduced in dimensionality by introducing a 1×1 kernel with weight multipliers before expansion. Most importantly, low-frequency environmental transformation basis is removed through forced difference operations. With high-frequency transformation matrix The formula for gating the output using the difference in feature distribution through a sigmoid monotonic activation function is as follows: ; in, The final output of the edge enhancement module is the enhanced feature map. This is the input feature map for the high-low frequency separation module. This is a two-dimensional average pooling operation. It is a sigmoid monotonic activation function. The convolution weight matrix is ​​the high-frequency feature branch. As a low-frequency environmental feature basis, it characterizes the background smoothing component of the feature map. As the bias vector of the high-frequency feature branch, this dual-channel stripping subtraction directly removes the aliasing noise source that causes interference at the lower boundary of the low signal-to-noise ratio without increasing the computational depth.

[0041] The feature deconstruction matrix of the constructed two-dimensional Haar wavelet frequency domain adaptive optimization layer is defined as follows: The module contains four sets of completely orthogonal, non-overlapping high-pass and low-pass operator spatial convolution kernels. The low-frequency global approximate energy operator is constrained to... The high-frequency sensitive operator for horizontal detail edges is limited to The vertical edge-sensitive operator is limited to The diagonal texture information operator is strictly limited to... .

[0042] Perform eigenvalue-level dimensionality expansion and recombination on the above four sets of basic operator tensors to establish a micro-dictionary structure matrix. By performing a product operation using a special subdivision sliding convolution with a fixed stride of 2, we obtain: ; in, This is the frequency domain feature map output by the Haar wavelet convolution. To complete the input feature map after boundary padding, For the Haar wavelet micro-dictionary structure tensor, For grouped convolution operations, the convolution stride s=2 and the number of channel groups is C. With extremely robust theoretical proof, it guarantees that the frequency domain decimation step achieves a nameless downsampling effect that completely eliminates spectral overlap and information distortion loss.

[0043] After introducing Haar wavelets, the network utilizes a feature calibrator to perform bi-domain feature interaction fusion, which includes: Representation matrix generated by spatial substructure layers With frequency domain basis response matrix After time synchronization and alignment, the merge function is executed. Injected into the fully connected coupled layer network. Subsequently, the network generates the fully connected factor tensor formula for controlling the nonlinear weights of importance within a very short time slot: ; in, This is the spatial domain representation feature matrix output by the spatial domain branch. The frequency domain basis response characteristic matrix of the frequency domain branch output. To calibrate the weight matrix of the convolutional layer, To calibrate the bias vector of the convolutional layer, For group normalization operation, The weighted layer calibrates the weight tensor for bi-domain features using a mathematically fully soft-assigned selection basis. This forces the network architecture to possess the core capability of continuously gliding and damping between the feature attention focus and the distribution pattern in the pure spatial domain or the edge of the high-frequency domain, based on the drastic fluctuations in the signal-to-noise ratio. This serves as the base feature map for the input of the fusion module, i.e., the original input features before the dual-domain branching process. This is the weighted matrix for the spatial channels.

[0044] In one embodiment, the training steps of the time-frequency signal parameter extraction network include: constructing a training sample set; the training sample set includes multiple sets of time-frequency image samples and signal component annotation information corresponding to each set of time-frequency image samples, the signal component annotation information including normalized bounding box coordinates converted from real time domain start and end information, center frequency and physical bandwidth; and training the pre-constructed time-frequency signal parameter extraction network using the training sample set to obtain the trained time-frequency signal parameter extraction network.

[0045] In one embodiment, the step of synchronously hard-truncating and slicing the multi-element observation matrix of each observation station according to the time-domain start and end information, and decoupling the hybrid baseband signal into multiple quasi-single target signal segments, includes: using the time-domain start and end information as a globally unified cutting mask, synchronously hard-truncating all multi-element observation matrices of each observation station in the time dimension; extracting data within the time interval corresponding to the time-domain start and end information to generate a local array observation matrix, and generating the corresponding quasi-single target signal segment based on the local array observation matrix.

[0046] Specifically, to completely eliminate the catastrophic coupling of feature space caused by traditional multi-source concurrency, the system abandons the traditional strategy of blindly searching the global matrix directly. Instead, it utilizes the time-domain start markers that the network autonomously detects and outputs with high fidelity. With termination mark This serves as a precise segmentation mask for the global data buffer. It applies to all distributed systems. Each base station, in the time dimension, captures the massive complex baseband observation data matrix of the original data. Simultaneous hard truncation slices are applied. After extraction and recombination, a sequence unique to the first slice is generated. Local array observation matrix of radiation events of individual targets This step mathematically reduces the solution of multi-target high-nonconvex hybrid spatial localization to a sequence of pseudo-single-target high-resolution solution operators in a time-slice dimension.

[0047] In one embodiment, the step of 3D spatial localization of a quasi-single target signal segment to obtain the corresponding preliminary 3D coordinates includes: calculating the spatial covariance matrix of each quasi-single target signal segment and performing eigenvalue decomposition to obtain the corresponding signal subspace and noise subspace; constructing a pseudo-inverse adaptive feature weighting matrix based on the signal subspace and an orthogonal projection matrix based on the noise subspace; calculating the weighted signal spatial projection energy of the pseudo-inverse adaptive feature weighting matrix at any 3D spatial point and the residual noise spatial projection energy of the orthogonal projection matrix at a 3D spatial point; constructing an enhanced feature spatial localization cost function based on the ratio of the weighted signal spatial projection energy to the residual noise spatial projection energy; and performing global optimization in the 3D search space using a pre-set Harris Eagle optimization method, and using the coordinates of the 3D spatial point that maximizes the enhanced feature spatial localization cost function as the preliminary 3D coordinates corresponding to the quasi-single target signal segment.

[0048] Specifically, the extracted locally truncated slice observation data matrix Traverse the discrete search coordinate vectors in three-dimensional space Based on the spatial geometric relative topological relationships between the system's transmitting and receiving nodes, the first... Array steering vector of each base station Subsequently, within the snapshot window of the current time-domain slice, the finite-sample local covariance matrix of the subarray is calculated. By forcibly applying eigenvalue decomposition to the matrix, the mixed signal space is forcibly separated in the orthogonal domain into a target signal subspace spanned by the preceding maximum eigenvalues ​​and their corresponding eigenvectors. And the additive white noise subspace spanned by the post-minimum eigenvalues ​​and their corresponding eigenvectors. The formula for obtaining the orthogonal projection matrix of its noise subspace is: ; in, Let be the orthogonal projection matrix of the noise subspace of the m-th base station. This is the conjugate transpose operation of a matrix. Let be the noise subspace matrix of the m-th base station. To overcome the bottleneck of pseudo-peak suppression and enhance the perception resolution of weak sidelobe targets in harsh environments with low signal-to-noise ratios, an Enhanced Feature Space (EES) fusion criterion is introduced. The formula for constructing the pseudo-inverse adaptive feature weighting matrix by extracting the singular value components of the signal subspace is as follows: ; in, It is a pseudo-inverse adaptive feature weighting matrix. Let m be the signal subspace matrix of the m-th base station. Let be the inverse of the diagonal matrix of the eigenvalues ​​of the signal subspace of the m-th base station. This weighted matrix mechanism utilizes the reciprocal property of the signal energy spectrum to adaptively reduce slippage and weaken the strong masking sidelobes generated by near-field high-power radiation sources in the radiation pattern. Subsequently, data-level incoherent diversity fusion is performed on the M distributed observation stations across physical space. By utilizing the quotient distribution of the weighted signal spatial projection energy and the residual noise spatial projection energy, a high-dimensional non-convex enhanced feature space localization cost function with strong signal resolution and noise floor suppression capabilities is jointly constructed. .

[0049] Its physical covariance theory foundation formula is established as follows: Under the physical assumption of infinitely approaching the limit of observation, the principle of complete orthogonality of classical subspaces requires... The condition holds true for all cases, therefore this constraint is proposed and the projection operator matrix is ​​minimized. The minimum value of different terms within the direction vector is taken. This differs from the classic, conventional approach of crudely constructing the formula by setting the denominator to the sum of residual energy. ; in, This is the spatial spectrum output of the classic MUSIC algorithm. Let p be the array steering vector corresponding to the m-th base station. Let be the target position coordinate vector to be estimated in three-dimensional space. To represent the total number of distributed observation base stations, a probabilistic weighted adjustment engine is introduced into the numerator. An adaptive reweighting matrix formula is constructed, which includes a diagonal inverse expansion of the principal large eigenvalues: ; in, The matrix pseudo-inverse operator, whose embedded negative first-order diagonal response, at the mathematical integral level, exerts absolute suppression on the high eigenvalue components contributed by extremely powerful interfering base stations, greatly flattening the energy gap. This allows even weak non-main-lobe small burst targets to break through the cover of environmental noise background and produce sharp spatial spectral peaks in the calculation of the following joint formula: ; in, The regularization constant is used for the mathematical derivation of the Cramer-Rao lower bound variance matrix of the theoretical limit in the system architecture. To verify the ultimate accuracy of the constructed operator, under stringent conditions such as continuous observation of complex Gaussian white noise and absolute independence of multiple sources, the objective unknowns are solved with respect to the three-dimensional divergence parameter. The derivation model of the centralized Fisher information matrix is ​​defined as follows: ; in, For position parameters Fisher's information matrix, This represents the number of snapshots (time sampling points). Let be the observation steering partial derivative matrix of the m-th base station and the k-th snapshot. The location corresponding to the m-th base station The orthogonal complement space projection operator, For operations involving the real part of complex numbers, the complement operation of projection onto the null space of the global signal subspace satisfies the following formula: ; in, Let L be an L-order identity matrix, where L is the number of antenna array elements in a single base station. Let m be the array manifold matrix of the m-th base station. As the inverse operation of the matrix, under the ideal condition of complete decoupling between the asymptotic limit and the channel, this heavy high-dimensional matrix rigorously collapses and decouples into... A pure 3×3 third-order independent partial derivative sub-block matrix: ; in, This is the 3×3 Fisher information block matrix corresponding to the q-th target. The signal power of the radiation source. Let be the partial derivative matrix of the guidance vector of the m-th base station with respect to the q-th target with respect to the three-dimensional position. The lower bound of the multi-target root mean square error is locked as . ; in, This is the lower bound of the Cramer-Rao formula for the root mean square error of 3D localization. For the trace operation of a matrix, The total number of radiation source targets is represented by the partial derivative tensor, which ensures extreme objectivity and rigor in the evaluation of the lower limit of convergence of the algorithm.

[0050] Due to the construction It exhibits strong nonconvexity and is highly prone to inducing local precocious puberty. The formula for minimizing the fitness evaluation index is defined as follows: ; An elite-guided lens imaging adaptive Harris Eagle optimization algorithm is introduced to drive the spatial solution engine. Within the population iterative evolution framework, an inverted S-shaped mathematical function is applied to control the prey's basic escape energy to undergo highly nonlinear iterative decay. When the population iterative evolution reaches the mid-to-late stage and spatial convergence stagnates, a dynamic lens imaging back-learning mechanism conforming to the laws of physical optics is immediately activated. Using the physical geometric core of the current search area as the virtual lens optical center, a dynamic scaling factor that exhibits a quadratic nonlinear convergence trend with increasing iteration frequency is introduced to force inferior solutions to perform large-scale back-jump projection in a wide-area symmetric space. After global exploration and local deep mining at the maximum iteration threshold, the coordinate vector that causes the final fitness to fall into a global minimum is forcibly output as the initial three-dimensional physical location point of the slice. .

[0051] In one specific implementation of the population decline heuristic control, the mathematical reconstruction process of the inverted S-shaped nonlinear energy regulation engine that affects the Harris Eagle's optimization step size and encircling oscillation radius includes: The extremely complex and smooth linear coefficient decay constant of the original Harris Eagle was abandoned, and a nonlinear inertial contraction damping base was constructed: ; in, To optimize the current iteration count of the algorithm, The maximum escape energy in the initial stage of the iteration. This represents the minimum escape energy at the end of the iteration. It is a monotonic function of inverted S-shape (Sigmoid type). To calculate the final real-time escape kinetic energy formula, which incorporates parabolic attenuation and sine wave interference terms to enhance the extreme sensitivity to complex electromagnetic micro-wave trough terrain, the base attenuation energy is adjusted to the inverted S-shaped control. ; in, For the final real-time escape kinetic energy, To optimize the maximum number of iterations of the algorithm, this nonlinear composite decay equation uses secondary decay in the later stage to strictly constrain the oscillation radius, while using a sinusoidal pulse waveform with continuously increasing frequency to ensure that the individual escapes the potential local optimum.

[0052] In one specific embodiment of the stop-and-go escape and dynamic virtual population leadership mechanism, the system's feedback response logic when spatial multiple extrema cause iterative population congestion includes: Configure a dedicated detection counter to monitor the speed of the spatial center of mass movement. When the population center of gravity has been quiescent for ten consecutive generations (i.e., When this occurs, the system forcibly triggers the suppression of computing power allocation for ordinary convergent individuals, thereby expanding the capacity of the elite exploration group and defining it as follows: ; in, For the first The number of elite populations in each generation To optimize the overall population size of the algorithm, during the normal convergence iteration phase when the population centroid movement rate is high, the size of the exploration group is adaptively reduced according to the following formula: ; At the same time, the top three seeds were forcibly extracted from the high-quality memory slots in the historical cache to construct the memory array. By mixing ratio coefficient Reconstructing the illusory and intangible formula of leadership beacon: ; in, For the merged leadership position, This represents the position of the historical global optimal solution that has been iterated to date. To determine the optimal individual position for the current generation, a discrete control factor that decays exponentially is used to regulate the step size of the population position iteration update. The discrete control factor is: ; in, The step size is used as a discrete control factor and as a step tolerance. The position coordinates are updated using a random coordinate weighted traction formula to overcome the attraction of local extrema and enhance the algorithm's ability to escape local optima.

[0053] In one specific example of optical physics lens imaging reverse projection variation, when the iteration process enters the middle and late stages (i.e. When this occurs, the dynamic reverse learning process includes: The system obtains the highest quality or most divergent index in the current space. As the original target image surface to be corrected, the central physical point calculated using the current full-dimensional boundary constraints is used as the virtual optical center lattice of the refractive glass. .in, and To avoid the infinite oscillation loop of conjugate beam loop caused by fixed focal length reflection, a soft focal length adjustable magnification transformation factor is established, where the objective lens shape shrinks drastically. ; in, As a soft-focal-adjustable magnification factor (lens magnification), based on the fundamental physical quantities of linear optical geometric symmetry projection, a reverse abrupt change landing point with forced reversal and scale compression is generated: ; in, The new individual position is generated by the lens inverse projection. If the fitness value of the new position obtained by the lens imaging inverse learning mechanism is better after substituting it into the cost function, the system will discard the original optimal coordinate and replace it with the new position.

[0054] In one embodiment, unsupervised spatial clustering based on spatial distance to obtain multiple clusters includes: setting a connectivity distance threshold based on the system positioning error; constructing an undirected graph based on the spatial distance and the connectivity distance threshold; extracting each connected component from the undirected graph; and treating each connected component as a cluster to complete the unsupervised spatial clustering process.

[0055] Specifically, after completing the isolated calculation of all component slices within a single reconnaissance window, the system obtains a discrete three-dimensional coordinate result including time series redundancy and spatial thermal noise distribution error. To achieve topological backtracking from isolated slices to continuous targets, the global space Euclidean distance scalar matrix of the set is calculated. Set a strict tolerance threshold for connectivity distance. A binary adjacency graph matrix of an undirected network graph model is constructed based on hard threshold decision logic. By running the connected component traversal extraction operator within the graph theory topological domain, all discrete points exhibiting spatial clustering are forcibly divided and merged into C sets of structurally independent and spatially disconnected clusters, without requiring prior input of the target number. This allows for a blind assessment of the number of real drone radiation sources and precise identification and attribution matching of sliced ​​entities.

[0056] In the process of constructing unsupervised geometric graph association maps, the core derivation process of the connection weight matrix for blind clustering of isolated slices across domains includes: The system extracts an absolutely scattered and disordered spatial estimate, i.e., a set of vectors, from two historical snapshots within a time window of arbitrary distribution. and The spatial distance rigidity is invoked to calculate the scalar difference coefficient between the computational kernel. : ; This leads to the construction of a global distance thermodynamic characterization matrix that reflects the density state of the connection coefficients between all global cutoff points. Extract the fixed rigid tolerance limit radius constant of the current deployment error limit constraint. This serves as a Boolean value trigger switch for edge weights. As long as the following condition is met... binary matrix Immediately marked as 1 by the deferred path; otherwise, completely disconnected and isolated, marked as 0. This zero-to-one transition strict graph theory Boolean decision network utilizes a breadth-first connectivity scanning mechanism to achieve autonomous identity archiving under the premise of no missing support for the number of bees and individual tracking labels.

[0057] In one embodiment, the step of statistically fusing the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum includes: performing centroid smoothing fusion processing on the three-dimensional coordinates of all signal segments within the same cluster to obtain the fused spatial coordinates of the radiation source; performing statistical averaging processing on the center frequency and physical bandwidth of all signal segments within the same cluster to obtain the fused electromagnetic parameters; and mapping the fused spatial coordinates and fused electromagnetic parameters corresponding to each radiation source to the three-dimensional geographic situation level to generate an electromagnetic spectrum.

[0058] Specifically, for the same identity cluster set in complex scenarios By analyzing data from multiple overlapping regions at different observation times, a deep, smooth parameter fusion across spatiotemporal physical dimensions is performed to completely eliminate abrupt errors from single-point measurements in accidental environments. In the three-dimensional physical coordinate system, the geometric relationships between the measurement sets of all hyperplane positioning nodes are obtained, and their high-precision final three-dimensional fusion core coordinates are synthesized. At the spatial level of radio electromagnetic parameters, the intrinsic frequency parameters of all signal slices collected from this cluster (including the center operating frequency term output by the network detection) are analyzed. With physical modulation bandwidth column By performing rigorous statistical weighted averaging and filtering to reduce noise, the ultimate electromagnetic physical properties of the target are obtained. and Ultimately, the system integrates and extracts the three-dimensional spatial base location. With frequency domain electromagnetic fundamental parameters Heterogeneous parameters, such as continuous survival time, are used to construct a full-dimensional situational feature fingerprint matrix for a single UAV through mapping theory. The entire system relies on high-precision, high-computing-power hardware for main control, and samples at specific time intervals. The complete strategy chain of the present invention is continuously tracked and triggered, outputting a dynamic electromagnetic situation panoramic view with time tracking and a global physical scale.

[0059] The effectiveness of the method of the present invention was verified through experiments, and a system was constructed using... A distributed positioning and monitoring system consisting of several observation stations, with each station deploying coordinates. They are respectively m. Each observation station is equipped with a uniform circular array consisting of L=8 elements, the array radius of which is determined by the actual physical emission frequency of the radiation source. The corresponding half-wavelength criterion is designed. The simulation environment is implemented based on the free-space path loss model, and the simulation scenarios cover... The simulation scenario involves a three-dimensional spatial region. It includes four UAVs flying in formation, radiating targets with initial real-world positions as follows: ; in, , , , The initial true 3D coordinate column vectors of the four UAV radiation sources in the simulation scenario are given. In the kinematic model, all aircraft move in a uniform straight line towards the center of the region at a speed of 10 m / s, with a constant Z-axis altitude during flight. The system triggers a positioning calculation every 10 seconds. At the signal level, each radiation source uses a TDMA mechanism with a transmission bandwidth of [missing information]. The broadband OFDM signal has a constant transmit power for each target. The receiver's local oscillator center frequency is set to... The system baseband adoption rate is The time window for a single positioning signal acquisition is set to The receiver noise floor power is constant. Regarding the parameter configuration of the localization algorithm, assuming a coarse grid search step size of 100m in 3D space, the search range of the ELA-HHO fine search stage is limited to the area around the coarsely estimated point. Within its neighborhood, the population size is set to The maximum number of iterations is set to 100. The connectivity distance threshold for the spatial clustering stage is set to... Taking the central monitoring station located at coordinates (2500, 2500, 0) as a typical observation node, TDMA network hybrid signals from four UAV radiation sources were collected. First, real-time analysis was performed, and the time consumption of each core step in a single processing flow was statistically analyzed, as shown in Table 1: Table 1. Time Consumption of Core Steps

[0060] The total time for a single map construction is 4.2867s, which is less than the system's set 10s positioning update cycle, thus meeting the real-time requirements for continuous dynamic tracking of radiation sources.

[0061] To verify the dynamic target tracking capability of the method of this invention in a continuous time-domain scenario, a simulation environment of dynamic electromagnetic situation mapping during UAV formation flight was constructed. The system performs signal acquisition and positioning fusion calculations with a 10-second cycle, displaying the system output situation at four key observation times: t=0s, 20s, 40s, and 60s. Figures 2 to 5 As shown.

[0062] In the figure, the pentagram marks the actual UAV trajectory, the dotted circle represents the estimated trajectory calculated and fused by the EES-ELA-HHO-DPD algorithm, and the square represents the initial position of each target. Simulation results show that as the observation time progresses, the cumulative estimated trajectory of each radiation source maintains a high degree of spatial consistency with the actual trajectory, and the overall relative geometric configuration of the formation is completely maintained throughout the 60-second tracking period.

[0063] To quantify the dynamic tracking accuracy of the system during continuous observation, the Euclidean distance between the estimated three-dimensional coordinates and the true coordinates of the target was used as the evaluation index for positioning error. Table 2 shows the detailed statistics of the real-time positioning error and the root mean square error over the entire observation process for each radiation source at seven observation steps. Table 2. Real-time positioning error and root mean square error of each radiation source at 7 observation steps.

[0064] The data in the table shows that during the 60-second continuous tracking process, the 3D positioning errors of the four targets remained stable within a range of 15m. Target 3 exhibited a maximum single estimation error of 27.09m at t=30s, but this did not trigger subsequent tracking divergence, and the system quickly converged its error to 2.13m at t=50s. The overall statistical results show that the RMSEs for the four targets were 7.80m, 7.00m, 11.90m, and 9.40m, respectively, with an overall average tracking RMSE of 9.03m.

[0065] In dynamic scenarios where the target's spatial location continuously changes, the system can stably complete time-domain signal detection, spatial parameter optimization, and unsupervised clustering tasks within each processing cycle. Experimental results confirm that the cascaded electromagnetic spectrum construction method proposed in this invention can achieve effective tracking of dynamic trajectories of multiple targets and accurate correlation of multi-dimensional electromagnetic features without requiring any prior information on the number of targets.

[0066] This invention constructs a cascaded spatiotemporal correlation framework of "slicing-localization-clustering," achieving deep integration between the front-end broadband spectrum sensing network and the back-end direct localization algorithm. By extracting time-frequency parameters, it performs time-domain adaptive slicing on mixed signals synchronously collected from multiple stations, transforming the complex problem of multi-source signal aliasing into an independent single-target sequential localization task. This fundamentally eliminates the dependence of traditional subspace algorithms on prior information about the number of radiation sources, enabling effective blind separation and precise decoupling of aliased signals from UAV swarms in unknown and complex electromagnetic environments. Simultaneously, it introduces an unsupervised spatial clustering method based on undirected graph connected components, autonomously constructing global spatial distances without manual intervention. The invention reliably classifies physical entities at discrete positioning points and blindly estimates the number of radiation sources. It also statistically smooths and fuses electromagnetic parameters such as the three-dimensional spatial position, center frequency, and signal bandwidth corresponding to the same radiation source, reconstructing fragmented spatiotemporal detection information into a complete and standardized multidimensional electromagnetic map. Furthermore, in engineering applications, this invention balances real-time processing and positioning accuracy. The single map construction process is short, meeting the requirements for high-frequency dynamic updates of electromagnetic situation. The overall three-dimensional positioning error is controllable, and the target motion estimation trajectory closely matches the actual trajectory. It can achieve continuous and stable tracking of multiple UAV radiation sources and accurate global electromagnetic situation perception in complex scenarios.

[0067] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0068] In one embodiment, an electromagnetic map construction device for UAV swarm detection is provided, comprising: The signal acquisition module is used to acquire the mixed baseband signal of the target UAV swarm, which is synchronously collected by multiple distributed observation stations; the mixed baseband signal includes the multi-element observation matrix of each observation station. The time-frequency transformation module is used to extract the one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image. The parameter extraction module is used to input the time-frequency image into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; The signal slicing module is used to perform synchronous time-domain hard truncation slicing of the multi-element observation matrix of each observation station according to the time-domain start and end information, and decouple the mixed baseband signal into multiple quasi-single target signal segments. The clustering module is used to perform three-dimensional spatial localization of a single target signal segment, obtain the corresponding preliminary three-dimensional coordinates, calculate the spatial distance between all preliminary three-dimensional coordinates, and perform unsupervised spatial clustering based on the spatial distance to obtain multiple clusters; each cluster corresponds to a radiation source. The spectrum generation module is used to statistically fuse the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and physical bandwidth.

[0069] Specific limitations regarding the electromagnetic map construction device for UAV swarm detection can be found in the limitations of the electromagnetic map construction method for UAV swarm detection described above, and will not be repeated here. Each module in the aforementioned electromagnetic map construction device for UAV swarm detection can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0070] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an electromagnetic spectrum construction method for UAV swarm detection. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0071] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does 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 those shown in the figure, or combine certain components, or have different component arrangements.

[0072] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0073] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0074] Those skilled in the art will understand that all or part of the processes in the methods of 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 of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] 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.

[0076] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing electromagnetic maps for UAV swarm detection, characterized in that, The method includes: Acquire the hybrid baseband signal of the target UAV swarm, which is synchronously collected by multiple distributed observation stations; the hybrid baseband signal includes the multi-element observation matrix of each observation station; Extract a one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image; The time-frequency image is input into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; Based on the time-domain start and end information, the multi-element observation matrix of each observation station is synchronously hard-truncated in the time domain and sliced ​​to decouple the mixed baseband signal into multiple quasi-single target signal segments. The quasi-single target signal segment is spatially located in three dimensions to obtain the corresponding preliminary three-dimensional coordinates. The spatial distance between all preliminary three-dimensional coordinates is calculated, and unsupervised spatial clustering is performed based on the spatial distance to obtain multiple clusters. Each cluster corresponds to a radiation source. The three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source are statistically fused to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and the physical bandwidth.

2. The method according to claim 1, characterized in that, The step of performing time-frequency transformation on the one-dimensional time-domain observation sequence and generating a time-frequency image includes: A short-time Fourier transform is performed on the one-dimensional time-domain observation sequence to obtain a two-dimensional time-frequency energy distribution; The two-dimensional time-frequency energy distribution is subjected to logarithmic compression and linear scaling to generate a standardized grayscale time-frequency image.

3. The method according to claim 1, characterized in that, The time-frequency signal parameter extraction network includes an edge-enhanced multi-scale feature extraction module and a frequency-space interaction detection head; The edge enhancement multi-scale feature extraction module extracts multi-scale features and separates high and low frequency components from the input time-frequency image, and outputs an enhanced multi-scale spatial feature map. The frequency-space interaction detection head performs orthogonal wavelet frequency domain lossless dimensionality reduction on the enhanced multi-scale spatial feature map to generate a frequency domain feature map. The spatial feature map and the frequency domain feature map are then fused together with dual-domain features and adaptive weight calibration is performed. Finally, the normalized bounding box coordinates of each signal component are output through bounding box regression.

4. The method according to claim 1, characterized in that, The training steps for the time-frequency signal parameter extraction network include: Construct a training sample set; the training sample set includes multiple sets of time-frequency image samples and signal component annotation information corresponding to each set of time-frequency image samples, the signal component annotation information includes normalized bounding box coordinates obtained by converting the real time domain start and end information, center frequency and physical bandwidth; The pre-constructed time-frequency signal parameter extraction network is trained using the training sample set to obtain the trained time-frequency signal parameter extraction network.

5. The method according to claim 1, characterized in that, The step of performing synchronous time-domain hard truncation slicing on the multi-element observation matrix of each observation station based on the time-domain start and end information, and decoupling the mixed baseband signal into multiple quasi-single-target signal segments, includes: Using the aforementioned time-domain start and end information as a globally unified cutting mask, all multi-element observation matrices of each observation station are synchronously hard-truncated in the time dimension. Data within the time interval corresponding to the start and end information in the time domain is extracted to generate a local array observation matrix, and a corresponding quasi-single target signal segment is generated based on the local array observation matrix.

6. The method according to claim 1, characterized in that, The step of performing three-dimensional spatial localization on the quasi-single-target signal segment to obtain the corresponding preliminary three-dimensional coordinates includes: Calculate the spatial covariance matrix of each quasi-single-target signal segment and perform eigenvalue decomposition to obtain the corresponding signal subspace and noise subspace; A pseudo-inverse adaptive feature weighting matrix is ​​constructed based on the signal subspace, and an orthogonal projection matrix is ​​constructed based on the noise subspace. Calculate the weighted signal spatial projection energy of the pseudo-inverse adaptive feature weighting matrix at any three-dimensional spatial point, and the residual noise spatial projection energy of the orthogonal projection matrix at the three-dimensional spatial point; An enhanced feature spatial localization cost function is constructed based on the ratio of the weighted signal spatial projection energy to the residual noise spatial projection energy. By using the pre-set Harris Eagle optimization method to perform global optimization in the three-dimensional search space, the coordinates of the three-dimensional spatial point that maximizes the enhanced feature spatial localization cost function are used as the initial three-dimensional coordinates corresponding to the quasi-single target signal segment.

7. The method according to claim 1, characterized in that, The unsupervised spatial clustering based on the spatial distance, resulting in multiple clusters, includes: A connectivity distance threshold is set based on the system positioning error, and an undirected graph is constructed based on the spatial distance and the connectivity distance threshold. Extract each connected component from the undirected graph and treat each connected component as a cluster to complete the unsupervised spatial clustering process.

8. The method according to claim 1, characterized in that, The step of statistically fusing the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum includes: Centroid smoothing fusion is performed on the three-dimensional coordinates of all signal segments within the same cluster to obtain the fused spatial coordinates of the radiation source. The center frequency and physical bandwidth of all signal segments within the same cluster are statistically averaged to obtain the radiative fusion electromagnetic parameters. The fused spatial coordinates and fused electromagnetic parameters of each radiation source are mapped to the three-dimensional geographic situation level to generate an electromagnetic map.

9. An electromagnetic map construction device for UAV swarm detection, characterized in that, The device includes: The signal acquisition module is used to acquire the hybrid baseband signal of the target UAV swarm simultaneously acquired by distributed multi-observation stations; the hybrid baseband signal includes the multi-element observation matrix of each observation station; The time-frequency transformation module is used to extract a one-dimensional time-domain observation sequence of any array element from the multi-element observation matrix, perform time-frequency transformation on the one-dimensional time-domain observation sequence, and generate a time-frequency image. The parameter extraction module is used to input the time-frequency image into a pre-trained time-frequency signal parameter extraction network to extract the time-domain start and end information, center frequency and physical bandwidth of each signal component; The signal slicing module is used to perform synchronous time-domain hard truncation slicing on the multi-element observation matrix of each observation station according to the time-domain start and end information, and decouple the mixed baseband signal into multiple quasi-single target signal segments. The clustering module is used to perform three-dimensional spatial localization on the quasi-single target signal segment, obtain the corresponding preliminary three-dimensional coordinates, calculate the spatial distance between all preliminary three-dimensional coordinates, and perform unsupervised spatial clustering based on the spatial distance to obtain multiple clusters; each cluster corresponds to a radiation source. The spectrum generation module is used to statistically fuse the three-dimensional coordinates and electromagnetic parameters of signal segments belonging to the same radiation source to generate an electromagnetic spectrum; the electromagnetic parameters include the center frequency and the physical bandwidth.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images