A method for locating a radiation source in an urban environment based on physical guidance and adaptive spatial focusing

Abstract

A kind of urban environment based on physical guiding and adaptive spatial focusing radiation source positioning method, this method first utilizes distributed receiving system to receive signal, constructs multimodal electromagnetic observation physical data model, generates multi-channel one-dimensional complex baseband sequence tensor and two-dimensional global space spectrum thermodynamic diagram;Subsequently, extract electromagnetic time domain and spatial geometric features through double-flow parallel topology structure, fuse and output global physical reference anchor point;Then affine transformation matrix is constructed with anchor point as center, and local continuous spatial feature map is obtained by resampling original high-resolution spatial energy spectrum;Further, extract local geometric features and output normalized geographic coordinate residual compensation vector;Finally, combined with reference anchor point, residual vector and physical scaling coefficient, calculate and output the final geographic coordinates of radiation source;The present application effectively solves the problems of poor robustness of traditional physical model and weak generalization ability of pure data-driven method in complex multipath environment in city, significantly improves the radiation source positioning accuracy and stability.

Description

Technical Field

[0001] This invention belongs to the field of passive positioning technology, specifically relating to a radiation source positioning method based on physical guidance and adaptive spatial focusing in an urban environment. Background Technology

[0002] With the rapid development of wireless communication technology and the widespread adoption of smart terminals, the rapid and accurate location of various non-cooperative radiation source targets (such as drones) in the increasingly complex urban electromagnetic environment is a critical task in fields such as radio spectrum management, urban security monitoring, and emergency rescue. However, the urban environment is characterized by high-density building distribution, irregular street layouts, and complex electromagnetic background noise. Electromagnetic waves inevitably experience severe non-line-of-sight (NLOS) transmission, multipath reflection, and diffraction during propagation. This complex physical propagation environment disrupts the ideal direct physical path, causing traditional passive positioning methods (such as Time Difference of Arrival (TDOA) and Angle of Arrival (AOA)) to face significant challenges due to the failure of the line-of-sight propagation assumption, resulting in a sharp increase in positioning errors.

[0003] In recent years, deep learning technology has been widely adopted in the field of passive localization due to its powerful nonlinear feature extraction capabilities. However, existing technologies still suffer from deep-seated physical and logical flaws in their processing approaches and model architectures, specifically in the following two aspects: Defect 1: The "visualization" of electromagnetic signals leads to the loss of core physical phase information and weak multipath resistance. Existing technologies, exemplified by Chinese patent application CN116299170A, "A Multi-Target Passive Localization Method, System, and Medium Based on Deep Learning," employ typical blind signal processing or image-like processing approaches. They typically transform one-dimensional baseband signals into two-dimensional time-frequency images or real-valued spectra, then use general-purpose visual convolutional networks for feature extraction. This purely visual processing paradigm severely disrupts the underlying physical logic of communication signals. First, the forced dimensionality increase consumes enormous computational resources; second, when generating two-dimensional real-valued images or performing frequency domain filtering and denoising, the crucial phase information in the original complex IQ signal is directly discarded. In distributed receiver systems, the phase difference of complex signals directly characterizes the absolute physical time delay and Doppler shift of electromagnetic waves during spatial propagation. General image processing networks cannot resolve this phase manifold based on the physical propagation of electromagnetic waves, which makes the model prone to false localization when faced with severe non-line-of-sight (NLOS) and multipath interference in urban canyons.

[0004] Defect 2: Traditional global-local visual frameworks cannot solve the bottlenecks of quantization errors and coordinate mapping in real continuous geographic space. To integrate prior environmental information, Chinese patent application CN120874583A, "Method and Device for Radiation Source Target Localization Based on Deep Learning in Urban Environments," introduces spatial rasterization technology, dividing the monitoring area into discrete grids. This type of method faces an irreconcilable contradiction between physical resolution and computational resources. To achieve high-precision localization while simultaneously conducting large-scale searches, the physical grid size must be reduced exponentially, leading to an explosive increase in network parameters and memory usage. More importantly, conventional approaches to resolving the global-local contradiction in existing technologies only involve pixel-level feature fusion and attention reweighting, which are completely incompatible with continuous regression tasks in physical geographic coordinate systems. Local refinement of image pixels does not involve real physical scaling of the geographic coordinate system; while radiation source localization requires a direct transition from the probability distribution of the wide-area spatial spectrum to the real continuous two-dimensional / three-dimensional geographic space. Existing technologies lack a mechanism that can achieve dynamic and differentiable spatial mapping and cropping in the actual physical coordinate system, making it difficult to smoothly achieve high-precision capture of large-scale searches and small positional perturbations in a single network. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, such as the loss of phase information in complex electromagnetic propagation environments and the inability of traditional static vision frameworks to solve the scaling and mapping of real physical coordinates, this invention provides a radiation source localization method based on physical guidance and adaptive spatial focus (ASF-Net) in urban environments. This method breaks the spatial discontinuity caused by fixed grid quantization and constructs a unified network architecture with dynamic scaling of physical field of view and feature concatenation, realizing end-to-end continuous high-precision mapping from underlying multimodal electromagnetic observation data to real target geographic coordinates.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A radiation source localization method based on physical guidance and adaptive spatial focusing in an urban environment, comprising the following steps: Step 1: Construct a multimodal electromagnetic observation physical data model, including signal modeling, data preprocessing, and physical-guided heatmap generation, which is the original high-resolution spatial energy spectrum; Step 2: Based on the multimodal electromagnetic observation physical data model, feedforward infers the wide-area reference anchor point. The reference coordinate extraction module with dual-stream parallel topology extracts and fuses features in the one-dimensional physical signal domain and the two-dimensional physical space domain respectively, and outputs the physical reference anchor point in the global real geographic coordinate system. Step 3: Achieve dynamic adaptive focusing of the physical field of view based on the real space mapping. Construct an affine transformation matrix centered on the physical reference anchor point within the focused field of view, and resample the two-dimensional global spatial spectrum heatmap to obtain a local continuous spatial feature map. Step 4: Complete continuous coordinate residual compensation within the focused field of view. Extract local geometric features through the local feature refinement and residual estimation module and output the normalized geographic coordinate residual compensation vector. Step 5: Reconstruct the end-to-end coordinate system. Combine the wide-area reference anchor point, residual compensation vector, and physical scaling factor to calculate and output the final geographic coordinates of the radiation source.

[0007] The signal modeling described in step 1 is as follows: For the distributed receiving system performing signal acquisition, the time-domain baseband signal received by the i-th receiving station in the system is modeled as a physical superposition of the direct path LoS component and multiple non-line-of-sight NLoS components. The expression is in the form of complex path gain, absolute physical propagation delay, Doppler frequency shift, number of virtual scatterers, and additive white Gaussian noise, as follows: in, For the first The first receiving station The complex path gain of the propagation path (where Corresponding to the direct trajectory Los component, (Corresponding to the non-line-of-sight NLoS component caused by the scatterer). This represents the absolute physical delay of propagation along the corresponding path described above. This refers to the Doppler frequency shift of the corresponding path mentioned above; The number of virtual scatterers in multipath propagation; For the first Additive white Gaussian noise superimposed from each receiving station; The data preprocessing involves discretizing and synchronously sampling the time-domain baseband signal output by the receiving system, stacking the time-domain baseband signals from all receiving stations into a multi-channel one-dimensional complex baseband sequence tensor, and losslessly preserving the absolute phase information representing the spatial physical distance. The receiving system is assumed to contain M receiving stations, with N discretized synchronous sampling points, and the discrete-time complex baseband sequence output by the i-th receiving station is... The multi-channel one-dimensional complex baseband sequence tensor The expression is: ; The physical guidance heatmap is generated as follows: using a spatial coherent accumulation algorithm based on the maximum eigenvalue, the monitoring area is discretized into a two-dimensional grid, and the matching response between the guidance vector of each receiving station and the covariance matrix of the received signal is calculated. The matching response calculation formula is as follows: Extract the maximum eigenvalue corresponding to each grid point. As the spatial spectral response intensity at that location, where the maximum eigenvalue physically characterizes the dominant electromagnetic signal energy at that spatial grid point, a two-dimensional global spatial spectral heatmap is generated in the real geographic coordinate system as a physical prior input.

[0008] Furthermore, the distributed receiving system includes multiple receiving stations, and its specific deployment rules are as follows: an outer circle is constructed with the physical geometric center of the target monitoring area as the reference position; five receiving stations are evenly deployed on the circumference of the outer circle according to a preset deployment radius, so that the angle between any two adjacent receiving stations and the reference position is 72°, thereby forming a spatial observation topology structure with a regular pentagonal layout.

[0009] The reference coordinate extraction module for the dual-stream parallel topology in step 2 includes a temporal-aware branch, a spatial topology branch, and an asymmetric fusion unit, specifically: The temporal-aware branch includes a signal quality-aware attention module and a one-dimensional convolutional backbone network. The signal quality-aware attention module performs global average pooling and max pooling along the time dimension to generate statistical descriptors for the pre-processed, stacked multi-channel complex baseband sequence tensor. It utilizes a bottleneck structure composed of dimensionality reduction, activation, and upscaling fully connected layers to learn the nonlinear dependencies between channels. Dynamic physical channel weights are generated through Sigmoid activation, and these weights are then multiplied channel-by-channel with the multi-channel complex baseband sequence tensor to achieve adaptive weighting. The enhanced sequence is input to a deep one-dimensional convolutional backbone network containing three cascaded convolutional blocks, and outputs an electromagnetic temporal feature vector through adaptive global average pooling. Spatial Topology Branch: The global spatial spectrum heatmap is downsampled using a deep convolutional encoder. A convolutional block attention module (CBAM) is introduced to suppress the electromagnetic background and focus on the region with the highest target energy using a spatial mask, mapping the global spatial spectrum heatmap into 512-dimensional spatial geometric features. An auxiliary decoder with a structure symmetrical to the encoder is set up. The 512-dimensional spatial geometric features are progressively upsampled using transposed convolution, and a noise-free ideal target binary mask is reconstructed. Physical consistency constraints are imposed by introducing mask reconstruction loss during the training phase, forcing the encoder to learn electromagnetic physical consistency and spatial geometric topology. Asymmetric fusion unit: It concatenates temporal features with spatial geometric features, inputs them into a three-layer fully connected layer with Mish activation, and outputs physical reference anchor points.

[0010] The specific process of dynamic adaptive focusing of the physical field of view in step 3 is as follows: taking the inferred physical reference anchor point as the center, setting the adaptive scaling ratio of the physical space, and dynamically constructing the affine transformation matrix; using the affine transformation matrix and the bilinear interpolation algorithm, directly performing grid resampling from the two-dimensional global spatial spectrum heat map, extracting and enlarging to obtain a local continuous spatial feature map of a fixed scale, and using the differentiable property of bilinear interpolation, the sampling grid of the local feature map propagates the gradient back to the physical reference anchor point, thereby maintaining the strict differentiable mapping relationship between the geometric space coordinates of the whole map and the local focused feature map.

[0011] The continuous coordinate residual compensation within the focused field of view described in step 4 specifically involves: performing lightweight CNN convolution processing on the resampled local continuous spatial feature map, extracting local physical high-frequency texture and peak morphology features, and flattening them to obtain a local geometric vector; inputting the local geometric vector into a regression network composed of multiple fully connected layers, and outputting a normalized geographic coordinate residual compensation vector relative to the center of the local physical field of view. The vector is constrained by the Tanh function to represent the actual physical distance perturbation.

[0012] The end-to-end coordinate system reconstruction described in step 5 is as follows: Substitute the wide-area reference anchor point, the normalized residual compensation vector, and the physical side length of the global monitoring area into the reconstruction formula: final geographic coordinates = wide-area reference anchor point + 0.5 × physical scaling factor × physical side length of global monitoring area × residual compensation vector, directly outputting the true geographic coordinates of the radiation source, realizing the end-to-end closed-loop mapping of electromagnetic observation data to the absolute physical coordinate system.

[0013] A radiation source localization system based on physical guidance and adaptive spatial focusing in an urban environment includes: The signal acquisition module is used to receive radiation source signals in the urban environment through a receiving system composed of multiple distributed single-antenna receiving stations; A processor for executing the radiation source localization method described above; The storage module is used to store urban building outline data, pre-trained adaptive spatial focusing neural network models, and physical propagation model parameters; The output module is used to display or transmit the calculated coordinates of the radiation source location.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the radiation source localization method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a joint network architecture that incorporates wide-area anchor point inference and local dynamic field-of-view scaling. By introducing a differentiable pruning mechanism based on affine transformation, a differentiable mapping between the local physical observation window and the global continuous geographic coordinate system is directly established within the network. This mechanism overcomes the inherent spatial quantization error in traditional rasterization methods, reduces the model's dependence on high-resolution grids under limited computing power, and achieves fine regression of real physical coordinates. This is fundamentally different from pixel-level visual feature refinement in the field of image processing.

[0016] For complex urban multipath environments, this invention forces the network to directly process one-dimensional complex baseband IQ sequences from multiple receiving stations. Since the phase of complex signals naturally encodes core geometric information such as the time difference of arrival (TDOA), which determines the absolute physical location of the target, and combined with an adaptive weighting mechanism for signal quality, the model can effectively resolve non-line-of-sight (NLOS) and multipath interference. This avoids the positioning performance loss caused by discarding physical phase in conventional real-valued or blind signal visualization methods, thus enhancing the robustness of the positioning system.

[0017] This method uses the spatial energy distribution generated by the spatial coherent accumulator array signal processing algorithm as a physical prior input, and introduces an auxiliary decoder with physical consistency constraints into the network topology branch. This mechanism, which integrates electromagnetic physical priors with deep features, guides the network to learn the real electromagnetic spatial geometry topology in the latent space, improving the problem of coordinate misjudgment that pure data-driven models are prone to when facing complex urban background noise and false spatial sidelobes. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of an embodiment of the present invention.

[0019] Figure 2 This is a diagram of the adaptive spatial focusing localization network architecture of the present invention, which combines physical guidance and multimodal fusion.

[0020] Figure 3 This is a probability distribution diagram from the global network perspective of this invention.

[0021] Figure 4 This is a schematic diagram of the physical space adaptive focusing process of the present invention.

[0022] Figure 5 This is a visualization of the positioning effect of the present invention.

[0023] Figure 6 This is a comparison chart of RMSE errors under different signal-to-noise ratios according to the present invention.

[0024] Figure 7 It is the cumulative distribution function (CDF) of positioning error under a specific signal-to-noise ratio in this invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0026] A physically guided adaptive spatial focusing and positioning method for unmanned aerial vehicles (UAVs) in an urban environment includes the following steps: Step 1: Construct a multimodal electromagnetic observation physical data model Signal Model: For a distributed receiving system performing signal acquisition, it is assumed that the system consists of 5 receiving stations arranged in a regular pentagonal pattern. This specific physical topology arrangement plays an irreplaceable role in achieving the high-precision positioning target of this invention in the following ways: (1) Isotropic geometric accuracy of spatial observation: The uniform central symmetric structure of the regular pentagon can provide a 360-degree spatial observation perspective without blind spots, so that the geometric accuracy factor (GDOP) in the monitoring area is minimized and evenly distributed, thus avoiding the blind spot problem of positioning error divergence in a specific direction from the physical level.

[0027] (2) Maximizing spatial diversity gain in multipath and non-line-of-sight (NLOS) environments: In densely built-up urban canyons, direct paths are easily blocked. Five distributed nodes construct rich redundant physical links, ensuring that at any target location, there is a high probability that at least some receiving stations can capture a reliable signal manifold. This provides sufficient link filtering margin for the "signal quality-aware attention module" in the network of this invention to extract high signal-to-noise ratio features.

[0028] (3) Suppressing false sidelobes of spatial coherent accumulation: The non-uniform distributed aperture formed by regular pentagons breaks the periodicity of conventional uniform grid sampling and can significantly suppress false grating lobes and spatial sidelobes in two-dimensional global spatial spectrum estimation, thereby providing a purer and unambiguous physical prior input for the subsequent "adaptive spatial focusing" process.

[0029] Based on the above system, the first in the system The time-domain baseband signal received by each receiving station is modeled as a physical superposition of the direct path (LoS) component and multiple non-line-of-sight (NLoS) components caused by environmental scatterers, specifically expressed as: (1) in, For the first The first receiving station The complex path gain of the propagation path (where Corresponding to the direct trajectory LosS component, (Corresponding to the non-line-of-sight NLoS component caused by the scatterer). This represents the absolute physical delay of propagation along the corresponding path described above. This refers to the Doppler frequency shift of the corresponding path mentioned above; The number of virtual scatterers in multipath propagation; For the first Additive white Gaussian noise superimposed from each receiving station. To reflect the actual propagation patterns of urban electromagnetic waves, the LoS path attenuation factor was set to 2.0, and the NLoS path attenuation factor was set to 3.8.

[0030] Data preprocessing: Discretize and synchronously sample the observed signals from the distributed receiving system, and then... The observation data from each receiving station are stacked into a multi-channel complex tensor. The complex form is retained here to preserve the absolute phase information (i.e., TDOA and FDOA parameters) that characterize the spatial physical distance without loss, completely abandoning the conventional technique of discarding phase and only taking amplitude as the input for the subsequent time domain branch.

[0031] Physically guided heatmap generation: Utilizing a spatial coherent accumulation algorithm based on the maximum eigenvalue, the monitoring area is discretized into a two-dimensional grid (a specific physical resolution can be set). The matching response between the receiving system's steering vector and the time-domain baseband signal covariance matrix is ​​calculated; the matching response calculation formula is as follows: Extract the maximum eigenvalue corresponding to each grid point. Since the maximum eigenvalue physically characterizes the dominant radiation source signal energy at this spatial grid point, it is used as the spatial spectral response intensity at that location. This allows for the calculation of the spatial electromagnetic energy distribution within the monitoring area in the real geographic coordinate system, generating a two-dimensional global spatial spectral heatmap. This heatmap represents the energy probability density of real space, rather than a simple visual image, and serves as the physical prior input for subsequent spatial topology branches. Step 2: Wide-area reference anchor point inference based on multimodal physical feedforward. A reference coordinate extraction module with a dual-stream parallel topology is constructed to extract electromagnetic time-domain features from the multi-channel one-dimensional complex baseband sequence tensor (i.e., the corresponding one-dimensional physical signal domain) and electromagnetic spatial geometric features from the two-dimensional global spatial spectrum heatmap (i.e., the corresponding two-dimensional physical spatial domain). These features are then fused through an asymmetric physical feedforward network to output physical reference anchor points in the global real geographic coordinate system.

[0032] Temporal-aware branch: includes a signal quality-aware attention module and a one-dimensional convolutional backbone network.

[0033] Signal quality awareness attention module: In order to suppress noise interference from severely flat fading channels at the physical level, this module performs global average pooling and max pooling on the input tensor along the time dimension to generate statistical descriptors, uses the bottleneck structure to generate dynamic physical channel weights, multiplies them element-wise with the original IQ input, and automatically filters out communication links containing reliable phases.

[0034] Deep temporal feature extraction: The enhanced sequence input contains a deep one-dimensional convolutional neural network with three cascaded convolutional blocks, and outputs an electromagnetic temporal feature vector through adaptive global average pooling.

[0035] Spatial topology branch: Extract topological features from the global spatial spectrum using a deep convolutional encoder.

[0036] Feature Encoding and Calibration: The input heatmap is downsampled by an encoder and a Convolutional Block Attention (CBAM) module is introduced to suppress electromagnetic background sidelobes and focus the main lobe energy. After calibration, it is mapped to 512-dimensional spatial geometric features.

[0037] Physical consistency constraint: Design a symmetric-assisted decoder. This module utilizes reconstruction loss to force the encoder to learn rigorous electromagnetic physical consistency and spatial geometric topology in the latent space, rather than image texture fitting, effectively overcoming the misjudgment of pure data-driven approaches when faced with spurious sidelobes.

[0038] Asymmetric fusion and benchmark anchor point output: Temporal features and spatial geometric features are concatenated, and through three fully connected layers (Mish activation), the physical benchmark anchor point of the target in the global real geographic coordinate system is output. .

[0039] Step 3: Dynamic adaptive focusing of the physical field of view based on real-space mapping To bridge the gap between wide-area search and local geometric constraints, a dynamic focusing mechanism based on physical affine transformation is directly introduced into the network.

[0040] Based on the inferred physical reference anchor point Centered on the physical space, set the adaptive scaling ratio to [value]. Dynamically construct affine transformation matrices within the network : (2) Using this matrix and the bilinear interpolation algorithm, grid resampling is performed directly from the original high-resolution spatial energy spectrum to extract and enlarge a fixed-scale local continuous spatial feature map. This process maintains a strictly differentiable mapping between the overall geometric spatial coordinates and the local focused feature map.

[0041] Step 4: Compensation for continuous coordinate residuals within the focused field of view The focused regression subnetwork is designed to eliminate nonlinear mapping biases caused by non-line-of-sight multipath propagation within the scaled physical field of view.

[0042] Local geometric feature extraction: Lightweight CNN convolution processing is performed on the resampled focused electromagnetic thermogram to extract the high-frequency fluctuation features of local spatial energy, and the local geometric vector is obtained after flattening.

[0043] Residual Spatial Migration Mapping: This method fuses feature inputs to a regression head and directly outputs a normalized geographic coordinate residual compensation vector relative to the local physical field of view center. (Constrained by the Tanh function). This vector represents the actual physical distance perturbation, rather than pixel offset.

[0044] Step 5: End-to-end coordinate system reconstruction Network terminals based on wide area reference anchor points and focused residual compensation vector Combined with physical scaling factor The final coordinates of non-cooperative targets are directly calculated and output in a continuous geographic coordinate system. : (3) in This represents the physical side length of the global monitoring area. This completes the end-to-end closed-loop mapping of electromagnetic observation data to the absolute physical coordinate system.

[0045] Example 1 Taking five known receiving stations and one non-cooperative drone target requiring localization in urban space as an example. The five receiving stations are arranged in a regular pentagon within the monitoring area, with the center coordinates as follows: The physical range of the monitoring area is The target moves randomly within the area. In this implementation method, all receiving stations are first synchronized in time and frequency, and a geometry-based random scattering body model (GSSM) is constructed to simulate the multipath electromagnetic propagation environment in the city.

[0046] This implementation process the one-dimensional complex baseband IQ sequence of multiple receiving stations, retains the core geometric information of the time difference of arrival (TDOA) and Doppler frequency shift (FDOA) encoded in the phase of the complex signal, and combines the signal quality adaptive weighting mechanism to analyze non-line-of-sight (NLOS) and multipath interference.

[0047] This implementation uses the spatial energy distribution generated by the spatial coherent accumulator array signal processing algorithm as a physical prior input, and integrates electromagnetic physical priors and deep features to guide the network to learn the real electromagnetic spatial geometric topology in the latent space.

[0048] The scenario state in this embodiment is as follows: Figure 1 As shown.

[0049] A radiation source localization method based on physical guidance and adaptive spatial focusing in urban environments is proposed. An adaptive spatial focusing network (ASF-Net) is constructed to achieve end-to-end continuous high-precision mapping from multimodal electromagnetic observation data to the true geographic coordinates of the radiation source. The specific steps are as follows: Step 1: Utilize the antenna to receive time-domain baseband signals, construct a multi-mode electromagnetic observation physical data model, perform signal modeling, data preprocessing, and physical-guided heatmap generation to obtain a multi-channel one-dimensional complex baseband sequence tensor and a two-dimensional global spatial spectrum heatmap. Specifically: The receiving stations are positioned in a two-dimensional space, and the number of receiving stations is initialized. Time and frequency synchronization of all receiving stations was determined. A geometry-based random scatterer model (GSSM) was constructed to simulate the urban multipath environment, setting the direct path (LoS) physical path attenuation factor to 2.0 and the non-line-of-sight (NLoS) scattering path attenuation factor to 3.8. The complex baseband signal received by each receiving station, which includes physical propagation delay and Doppler frequency shift, is modeled as follows: in, For the first The first receiving station The complex path gain of the propagation path (where Corresponding to the direct trajectory LosS component, (Corresponding to the non-line-of-sight NLoS component caused by the scatterer). This represents the absolute physical delay of propagation along the corresponding path described above. This refers to the Doppler frequency shift of the corresponding path mentioned above; The number of virtual scatterers in multipath propagation; For the first Additive white Gaussian noise is superimposed at each receiving station. The sampling signal length at each receiving station is determined to be 4096 sampling points.

[0050] Step 2: Five receiving stations simultaneously receive signals radiated from the radiation source target. After synchronization and cyclic prefix removal, the discrete sequence is truncated, and the real and imaginary parts are separated and stacked into a multi-channel one-dimensional complex baseband sequence tensor. The complex observation signal is processed using a spatial coherent accumulation algorithm based on the maximum eigenvalue, and the monitoring area is discretized into... Grid (e.g.) The maximum eigenvalue corresponding to each grid point is calculated as the spatial spectral response intensity. Then, Min-Max normalization is used to linearly map the pixel intensity to the [0,1] interval, forming a single-channel two-dimensional global spatial spectral heatmap. This heatmap explicitly incorporates prior physical and geometric information as input for the subsequent spatial vision branch.

[0051] Step 3: The reference coordinate extraction module of the dual-stream parallel topology extracts electromagnetic time-domain features from the multi-channel one-dimensional complex baseband sequence tensor, extracts electromagnetic spatial geometric features from the two-dimensional global spatial spectrum heatmap, and fuses and outputs physical reference anchor points in the global real geographic coordinate system. Specifically: Multimodal physical observation data is input into the reference coordinate extraction module of the dual-stream parallel topology for processing. First, the multi-channel complex baseband sequence tensor is adaptively weighted using a signal quality-aware attention module. Let the input complex baseband signal tensor be... ( The total number of channels. (where the signal length is 0), perform global average pooling and global max pooling along the time dimension to generate statistical descriptors. and : The merged descriptor Input to include a dimension reduction matrix and the increasing dimension matrix The bottleneck structure generates a dynamic physical channel weight vector. : in, It is the ReLU activation function. This is the Sigmoid activation function. Ultimately, it passes... Adaptive weighting is achieved by multiplying the inputs of each channel element by element, thereby effectively extracting high signal-to-noise ratio link features and suppressing noise interference; Simultaneously, a spatial vision branch is used to encode the 2D global spatial spectrum heatmap, and a convolutional block attention module (CBAM) is introduced to suppress background noise. Furthermore, a physically consistent auxiliary decoder symmetric to the encoder is constructed, and a multi-task joint loss function is introduced during the training phase. ,in For coordinate regression loss, To constrain the Focal Loss segmentation loss for reconstructing the topological morphology of the spatial spectrogram, the encoder is forced to learn the electromagnetic physical consistency and spatial geometric topology in the latent space, thereby extracting the topological features of the real geographic space. Finally, by splicing and fusing temporal complex features and spatial geometric features in the latent space, and then passing the result through a regression network using the Mish activation function, the normalized physical reference anchor point of the target in the geographic coordinate system is output. .like Figure 3 As shown, this is the probability distribution from a global perspective, demonstrating the network's performance on the ground. Figure 4 The global view probability distribution was inferred under test scenarios with different target locations (sample numbers: 9210, 3647, 4535, 6749). Within a 5000×5000 meter two-dimensional urban monitoring area, the normalized spatial energy response intensity was calculated using a spatial coherent accumulation algorithm. The bright centers where rays intersect in the figure reflect the spatial energy probability density distribution generated by the network based on physical guidance, providing a basis for subsequent inference of benchmark anchor points.

[0052] Step 4: Construct an affine transformation matrix centered on the physical reference anchor point within the focused field of view, and resample the original high-resolution spatial energy spectrum to obtain a locally continuous spatial feature map. Specifically: Perform a differentiable dynamic focus clipping operation in the physical coordinate system. The normalized physical reference anchor point is inferred from the reference coordinate extraction module. Centered on the local window, set the physical scaling ratio of the local window relative to the global scene to [value]. Construct an affine transformation matrix for the real geographic coordinate system. : Using this matrix and the bilinear interpolation algorithm, the original two-dimensional global spatial spectrum heatmap was obtained. Resampling is then performed. Specifically, a normalized sampling grid is generated. And through the formula Differentiable mesh sampling is performed, and while preserving gradient differentiability (allowing the supervision signal from the local refinement module to backpropagate the gradient to the global network), the sample is truncated and enlarged to obtain a size of... Local continuous spatial feature map ,like Figure 4 As shown, it was disassembled in detail. Figure 3 The dynamic focusing and error compensation process of the four test samples is shown. The left sub-figure shows the local focusing field of view constructed with the reference anchor point initially inferred by the network as the center; the middle column transforms the focusing area into a local continuous spatial feature map through affine transformation and bilinear interpolation algorithm, clearly presenting the microscopic geometric relationship between the real target position, the inferred reference point and the local multipath texture, reflecting the system's ability to transition from global search to local observation; the residual compensation vector diagram on the right intuitively indicates the normalized residual compensation vector calculated by the network regression module, which is used to compensate for the positioning deviation caused by non-line-of-sight and multipath propagation.

[0053] Step 5: Extract local geometric features from the local continuous spatial feature map using the local feature refinement and residual estimation module, and output a normalized geographic coordinate residual compensation vector. Specifically: The resampled local continuous spatial feature map is input into the local feature refinement and residual estimation module. This module uses lightweight feature extraction and nonlinear mapping structure to extract microscopic high-frequency textures, and outputs a normalized geographic coordinate residual compensation vector of the target relative to the center of the local physical window through a fully connected regression head. .

[0054] Step 6: Combining the wide-area reference anchor point, residual compensation vector, and physical scaling factor, calculate and output the final geographic coordinates of the radiation source. Specifically: Based on normalized physical reference anchor points and physical scaling factor Combined with the global physical side length of the monitoring area By substituting the parameters into the reconstruction formula, the final geographic coordinates of the radiation source are calculated and output directly in the continuous geographic coordinate system. (Unit: meters): like Figure 5 As shown, this is the final positioning result after residual compensation, and the predicted geographical location highly coincides with the actual location. Figure 5 The final geolocation results after end-to-end reconstruction are presented, quantifying the improvement in positioning accuracy achieved by the residual compensation module. In samples 9210, 3647, 4535, and 6749, the initial prediction errors of the reference anchor points were 62.6 m, 16.1 m, 84.7 m, and 59.0 m, respectively. After adaptive residual compensation by the system, the final positioning errors were reduced to 3.4 m, 1.2 m, 3.7 m, and 1.1 m, respectively. Experimental results show that in macroscopic urban monitoring areas, this network can effectively overcome disturbances caused by multipath propagation, controlling the positioning error within the range of 1.1 m to 3.7 m, verifying the effectiveness of the proposed adaptive spatial focusing network in achieving high-precision positioning.

[0055] Comparative experiments and analysis: Under the same physical simulation environment, the positioning performance of the method of this invention, the traditional TDOA positioning method, the traditional direct localization algorithm (DPD), and the Cramer-Rao theory lower bound (CRLB) were compared under different signal-to-noise ratios. The positioning estimation error is defined as the root mean square error of the Euclidean distance between the actual physical location of the target and the estimated location.

[0056] Depend on Figure 6As can be seen, within the signal-to-noise ratio (SNR) range of -10dB to 10dB, the RMSE of this invention is consistently lower than that of the traditional DPD algorithm. Under low SNR and multipath conditions (-10dB), the error of the traditional DPD algorithm increases sharply; however, this invention, by processing the phase difference of complex IQ signals and fusing electromagnetic physical priors, maintains a lower level of physical error and approaches the theoretical CRLB lower bound as the SNR increases. This reflects the effectiveness of the model in extracting underlying electromagnetic features in resisting multipath interference.

[0057] Depend on Figure 7 As can be seen from the CDF curves, at a signal-to-noise ratio of 0 dB, the curve of the method of this invention is significantly shifted to the upper left compared to the result using only global inference. This indicates that by introducing physical space dynamic focusing and residual compensation based on affine transformation, the long-tailed positioning error caused by multipath reflection is effectively eliminated, and the positioning error of more than 90% of the test samples is controlled within the meter range.

[0058] Experimental results show that this network architecture reduces the global high-resolution spatial grid memory requirements while improving the target positioning accuracy and robustness in complex multipath urban environments through continuous physical mapping in the real geographic coordinate system.

Claims

1. A radiation source localization method based on physical guidance and adaptive spatial focusing in an urban environment, characterized in that, Perform the following steps in sequence: Step 1: Receive time-domain baseband signals using a receiving system composed of multiple distributed single-antenna receiving stations, construct a multi-mode electromagnetic observation physical data model, perform signal modeling, data preprocessing, and physical-guided heatmap generation to obtain a multi-channel one-dimensional complex baseband sequence tensor and a two-dimensional global spatial spectrum heatmap; Step 2: Extract electromagnetic time-domain features from multi-channel one-dimensional complex baseband sequence tensors using the reference coordinate extraction module of dual-stream parallel topology, extract electromagnetic spatial geometric features from two-dimensional global spatial spectrum heat map, and fuse and output physical reference anchor points under the global real geographic coordinate system. Step 3: Construct an affine transformation matrix centered on the physical reference anchor point within the focused field of view, and resample the original high-resolution spatial energy spectrum to obtain a local continuous spatial feature map; Step 4: Extract local geometric features from the local continuous spatial feature map through the local feature refinement and residual estimation module and output the normalized geographic coordinate residual compensation vector; Step 5: Combine the wide-area reference anchor point, residual compensation vector, and physical scaling factor to calculate and output the final geographic coordinates of the radiation source.

2. The method according to claim 1, characterized in that, The signal modeling described in step 1 is as follows: For the distributed receiving system performing signal acquisition, the time-domain baseband signal received by the i-th receiving station in the system is modeled as a physical superposition of the direct path LoS component and multiple non-line-of-sight NLoS components. The expression is in the form of complex path gain, absolute physical propagation delay, Doppler frequency shift, number of virtual scatterers, and additive white Gaussian noise, as follows: (1) in, For the first The first receiving station Complex path gain of a propagation path Corresponding to the direct trajectory Los component, Corresponding to the non-line-of-sight NLoS component caused by the scatterer; This represents the absolute physical delay of propagation along the corresponding path described above. This refers to the Doppler frequency shift of the corresponding path mentioned above; The number of virtual scatterers in multipath propagation; For the first Additive white Gaussian noise superimposed from multiple receiving stations; The data preprocessing involves discretizing and synchronously sampling the time-domain baseband signal output by the receiving system, stacking the time-domain baseband signals from all receiving stations into a multi-channel one-dimensional complex baseband sequence tensor, and losslessly preserving the absolute phase information representing the spatial physical distance. The receiving system is assumed to contain M receiving stations, with N discretized synchronous sampling points, and the discrete-time complex baseband sequence output by the i-th receiving station is... The multi-channel one-dimensional complex baseband sequence tensor The expression is: . The physical guidance heatmap is generated as follows: using a spatial coherent accumulation algorithm based on the maximum eigenvalue, the monitoring area is discretized into a two-dimensional grid, and the matching response between the guidance vector of each receiving station and the covariance matrix of the received signal is calculated. The matching response calculation formula is as follows: Extract the maximum eigenvalue corresponding to each grid point. As the spatial spectral response intensity at that location, where the maximum eigenvalue represents the dominant radiation source signal energy at that spatial grid point, a two-dimensional global spatial spectral heatmap is generated in the real geographic coordinate system as a physical prior input.

3. The method according to claim 1, characterized in that, The reference coordinate extraction module for the dual-stream parallel topology in step 2 includes a temporal-aware branch, a spatial topology branch, and an asymmetric fusion unit, specifically: The temporal-aware branch includes a signal quality-aware attention module and a one-dimensional convolutional backbone network. The signal quality-aware attention module performs global average pooling and max pooling along the time dimension to generate statistical descriptors for the multi-channel one-dimensional complex baseband sequence tensor formed by stacking after preprocessing. It learns the nonlinear dependencies between channels using a bottleneck structure composed of dimensionality reduction, activation, and dimensionality increase fully connected layers, and generates dynamic physical channel weights through Sigmoid activation. The weights are then multiplied channel by channel by channel to achieve adaptive weighting. The enhanced sequence is input into a deep one-dimensional convolutional backbone network containing three cascaded convolutional blocks, and electromagnetic temporal features are output through adaptive global average pooling. Spatial topology branch: The two-dimensional global spatial spectrum heatmap is downsampled using a deep convolutional encoder, and a convolutional block attention module is introduced to suppress the electromagnetic background and focus on the region with the highest target energy through a spatial mask, thus mapping the two-dimensional global spatial spectrum heatmap into 512-dimensional spatial geometric features; An auxiliary decoder with a structure symmetrical to the encoder is set up. The 512-dimensional spatial geometric features are progressively upsampled using transposed convolution, and a noise-free ideal target binary mask is reconstructed. By introducing mask reconstruction loss during the training phase, physical consistency constraints are imposed, forcing the encoder to learn electromagnetic physical consistency and spatial geometric topology, and finally obtaining electromagnetic spatial geometric features. Asymmetric fusion unit: It splices electromagnetic temporal features with spatial geometric features, inputs them to a three-layer fully connected layer activated by Mish, and outputs physical reference anchor points.

4. The method according to claim 1, characterized in that, The specific implementation process of the physical field of view dynamic adaptive focusing in step 3 is as follows: taking the physical reference anchor point as the center, setting the adaptive scaling ratio of the physical space, and dynamically constructing an affine transformation matrix in the adaptive spatial focusing network; using the affine transformation matrix and the bilinear interpolation algorithm, directly performing grid resampling from the two-dimensional global spatial spectrum heat map, extracting and enlarging to obtain a local continuous spatial feature map of a fixed scale, and using the differentiable property of bilinear interpolation, making the sampling grid of the local feature map propagate the gradient back to the physical reference anchor point, thereby maintaining the strict differentiable mapping relationship between the geometric space coordinates of the whole map and the local focusing feature map.

5. The method according to claim 1, characterized in that, The continuous coordinate residual compensation within the focused field of view described in step 4 specifically involves: performing lightweight CNN convolution processing on the resampled local continuous spatial feature map to extract local physical high-frequency texture and peak morphology features and flattening them to obtain local geometric features; inputting the local geometric features into a regression network composed of multiple fully connected layers to output a normalized geographic coordinate residual compensation vector relative to the center of the local physical field of view. This vector is constrained by the Tanh function to represent the actual physical distance perturbation.

6. The method according to claim 1, characterized in that, The end-to-end coordinate system reconstruction described in step 5 is as follows: Substitute the wide-area reference anchor point, the normalized residual compensation vector, and the physical side length of the global monitoring area into the reconstruction formula: final geographic coordinates = wide-area reference anchor point + 0.5 × physical scaling factor × physical side length of global monitoring area × residual compensation vector, directly outputting the true geographic coordinates of the radiation source, realizing the end-to-end closed-loop mapping of electromagnetic observation data to the absolute physical coordinate system.

7. A radiation source localization system based on physical guidance and adaptive spatial focusing in an urban environment, characterized in that, include: The signal acquisition module is used to receive radiation source signals in the urban environment through a receiving system composed of multiple distributed single-antenna receiving stations; A processor for performing the radiation source localization method as described in any one of claims 1 to 6; The storage module is used to store urban building outline data, pre-trained adaptive spatial focusing neural network models, and physical propagation model parameters; The output module is used to display or transmit the calculated coordinates of the radiation source location.

8. 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 radiation source localization method as described in any one of claims 1 to 6.

Images