Radio map reconstruction method and device based on double diffusion model

Abstract

The invention discloses a radio map reconstruction method and device based on a double diffusion model, and relates to the technical field of communication, and the method comprises the steps: obtaining to-be-processed data; the trained first-stage network model is adopted to process to-be-processed data, spatial features of rapid change of signals are captured, and an electromagnetic singular region is obtained through prediction; processing the electromagnetic singular region and data to be processed by adopting the trained second-stage network model, and generating a radio map by taking the electromagnetic singular region as structural guidance; wherein the trained first-stage network model is obtained by taking a first sample as training data, taking a pre-constructed electromagnetic singular point contour map as a label and training an initial diffusion model, and the trained second-stage network model is obtained by taking a second sample as training data and taking a real radio map as a label. And training the initial diffusion model. According to the method, the map-based positioning and navigation capability can be remarkably enhanced.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a radio map reconstruction method and apparatus based on a dual diffusion model. Background Technology

[0002] With the development of 6G wireless communication, communication systems are gradually transforming into environment-aware systems, requiring network nodes to adapt to complex and ever-changing wireless environments in real time. Against this backdrop, Radio Maps (RMs) have become a key tool, supporting low-overhead channel estimation, UAV trajectory planning, and network optimization.

[0003] Existing radio mapping methods mainly fall into two categories: one is high-precision methods based on electromagnetic (EM) theory. These methods obtain accurate radio map reconstructive (RM) data by solving Maxwell's equations or using full-wave numerical simulation to calculate the spatial distribution characteristics of the radio channel, and can reflect the spatial variation patterns of the channel. While these methods can generate high-precision RM data that considers multipath effects, they are computationally intensive, typically assume a static environment, and are difficult to adapt to the real-time requirements of mobile network elements and dynamic scenarios in 6G networks.

[0004] Another category is data-driven methods based on neural networks (NNs). These methods rapidly predict radio maps by training models, significantly improving generation efficiency and enabling inference using limited measurement data. Existing NN methods are mostly used for main path signal modeling, achieving fast approximate RM generation, and are used in applications such as low-overhead channel estimation, mobile network optimization, and UAV trajectory planning. However, existing models typically only focus on the main path propagation characteristics, making it difficult to accurately capture signal abrupt changes and electromagnetic singularities caused by multipath propagation. This results in unclear boundaries and missing spatial details in reconstructed radio maps under complex environments, and fails to fully reflect high-frequency local variations caused by reflection, diffraction, and scattering. This limits their accuracy and reliability in practical applications such as UAV trajectory planning, low-overhead channel estimation, and dynamic network optimization.

[0005] Therefore, existing technologies are insufficient to simultaneously meet the comprehensive requirements of high precision, multi-path perception, and real-time generation. There is an urgent need for a method that can efficiently generate RM and accurately model multi-paths and singularities to meet the high-precision environmental perception requirements of future 6G networks. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a radio map reconstruction method and apparatus based on a dual-diffusion model. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a radio map reconstruction method based on a dual-diffusion model, comprising: Acquire the data to be processed; the data to be processed includes environmental information, building distribution, and base station locations; A pre-trained first-stage network model is used to process the data to be processed, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities. A pre-trained second-stage network model is then used to process both the electromagnetic singularities and the data to be processed, using the electromagnetic singularities as a structural guide to generate a radio map. The first-stage network model, trained using the first sample as training data and a pre-constructed electromagnetic singularity contour map as labels, is used to train the initial diffusion model. The second-stage network model, trained using the second sample as training data and a real radio map as labels, is used to train the initial diffusion model.

[0007] Secondly, the present invention also provides a radio map reconstruction apparatus based on a dual-diffusion model, comprising: The data acquisition module is used to acquire data to be processed; the data to be processed includes environmental information, building distribution, and base station locations. The data processing module uses a pre-trained first-stage network model to process the data to be processed, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities. A pre-trained second-stage network model then processes both the electromagnetic singularities and the data to be processed, using the electromagnetic singularities as a structural guide to generate a radio map. The first-stage network model, trained using the first sample as training data and a pre-constructed electromagnetic singularity contour map as labels, is used to train the initial diffusion model. The second-stage network model, trained using the second sample as training data and a real radio map as labels, is used to train the initial diffusion model.

[0008] The beneficial effects of this invention are: This invention provides a radio map reconstruction method and apparatus based on a dual diffusion model. By introducing a diffusion model guided by physical information, it achieves high-precision reconstruction and detail preservation of radio maps. It can effectively capture electromagnetic singularity structures and maintain stable performance in complex and dynamic wireless environments. This not only improves the reliability and usability of map reconstruction, but also significantly enhances map-based positioning and navigation capabilities, providing strong technical support for UAV path planning, intelligent transportation, and future wireless communication systems.

[0009] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0010] Figure 1This is a flowchart of a radio map reconstruction method based on a dual diffusion model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the training of the first-stage network model and the second-stage network model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of training a first-stage network model or a second-stage network model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the comparison of various indicators provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating a comparison in an SRM radio map reconstruction task provided in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating a comparison in a DRM radio map reconstruction task provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating a comparison in an MRM radio map reconstruction task provided in an embodiment of the present invention; Figure 8 This is a schematic diagram comparing the positioning errors of different methods provided in the embodiments of the present invention; Figure 9 This is a schematic diagram comparing the inference time and memory usage of different models provided in this embodiment of the invention; Figure 10 This invention provides an embodiment that uses different contour maps as... A diagram illustrating input performance comparison; Figure 11 This is based on the embodiments of the present invention. A schematic diagram of quantitative comparison on a map; Figure 12 This is provided by the embodiments of the present invention. A schematic diagram of a qualitative comparative map. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0012] This invention aims to address the challenge of balancing efficiency and accuracy in existing radio map construction methods in complex wireless environments. While existing high-precision methods based on electromagnetic theory can provide relatively accurate channel spatial distributions, they typically involve large computational loads and assume a static environment, making them unsuitable for the dynamic and ever-changing demands of 6G networks. Existing data-driven methods based on neural networks, although possessing rapid prediction capabilities, usually only focus on the main path propagation characteristics, failing to accurately capture electromagnetic singularities such as multipath effects, diffraction, and scattering, resulting in insufficient accuracy in radio map reconstruction under complex environments. To address this technical problem, this invention proposes a diffusion model-based radio map reconstruction method guided by physical information. By explicitly integrating electromagnetic singularity information into the diffusion model generation process, it effectively enhances the model's ability to perceive multipath effects and local signal abrupt changes.

[0013] Please see Figure 1 , Figure 1 This is a flowchart of a radio map reconstruction method based on a dual-diffusion model provided in an embodiment of the present invention. The radio map reconstruction method based on a dual-diffusion model provided by the present invention includes: S101. Obtain the data to be processed; the data to be processed includes environmental information, building distribution and base station location.

[0014] S102. The trained first-stage network model is used to process the data to be processed, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities. The trained second-stage network model is then used to process both the electromagnetic singularities and the data to be processed, using the electromagnetic singularities as a structural guide to generate a radio map. The first-stage network model, trained using the first sample as training data and a pre-constructed electromagnetic singularity contour map as labels, is used to train the initial diffusion model. The second-stage network model, trained using the second sample as training data and a real radio map as labels, is used to train the initial diffusion model.

[0015] Specifically, in this embodiment, please refer to Figure 2 and Figure 3 , Figure 2 This is a schematic diagram illustrating the training of the first-stage network model and the second-stage network model provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of training a first-stage network model or a second-stage network model provided in an embodiment of the present invention. The architecture of the trained first-stage network model is the same as that of the trained second-stage network model, both including: a variational autoencoder, a diffusion model, and a variational autodecoder; wherein, the diffusion model is a conditional denoising diffusion model. Variational autoencoders are used to convert input data Mapping to a lower-dimensional space yields the latent representation. And add noise to the latent representation. ; Diffusion model used for latent representation of superimposed noise The process involves modeling in the latent space, preserving the structural features of electromagnetic singularities and high-curvature regions through latent representation, and obtaining the predicted drift term. and predicted noise term Based on the predicted drift term and the predicted noise term, the reconstructed latent representation is obtained. , is represented as: ; Variational autodecoders are used to reconstruct the latent representation and obtain output data. .

[0016] The present invention proposes The framework employs a two-stage diffusion model for high-fidelity radio map generation. The first stage of the diffusion model predicts electromagnetic singularities, namely... The distribution of the area captures the spatial characteristics of rapidly changing signals; the second-stage model uses the output of the first stage as a structural guide, and combines environmental information and base station locations to generate the final radio map.

[0017] To improve computational efficiency and preserve spatial structure information, the framework proposed in this invention introduces a variational autoencoder (VAE) to perform latent spatial encoding on the radio map. The original radio map is first mapped to a low-dimensional latent representation through the VAE encoder. This representation retains the main spatial features and structural information while significantly reducing the data dimensionality. Subsequently, a diffusion model is used to model the latent space, preserving the structural features of electromagnetic singularities and high-curvature regions through the latent representation. Finally, the VAE decoder reconstructs the latent representation back to the original space, achieving high-fidelity radio map generation while accurately recovering spatial details and high-frequency signal variations.

[0018] In this embodiment, the trained first-stage network model is obtained by training the initial diffusion model using the first sample as training data and the pre-constructed electromagnetic singularity contour map as labels, including: Multiple first samples are obtained to construct a training dataset; where the first sample is a radio map including environmental information, building distribution and base station location; For radio maps, an electromagnetic singularity contour map is constructed as the true label for the first sample; Input a portion of the samples from the training dataset into the first... The first-stage network model to be trained is then trained to obtain the second... The prediction results output during this training process; where the prediction results include the predicted drift term. Predicted noise term The latent representation of reconstruction and output data ; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the first samples of the first-stage network model to be trained are used to calculate the loss and serve as the basis for the second-stage training. The loss during the training process; among which, the loss of the first training process; The loss in this training process includes the loss of the variational autoencoder decoder. Loss of the diffusion model The loss in the diffusion model includes the loss of the drift term. Loss of noise term and the loss of reconstructing the latent representation ; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the first-stage network model to be trained are obtained. The first-stage network model to be trained is then iterated until the number of training iterations or the degree of convergence meets the preset conditions, thus obtaining the trained first-stage network model.

[0019] Understandably, this applies to the first-stage network model. The training parameters are Used to determine environmental information Building distribution and base station location Predicted distribution map of electromagnetic singularities This singular region map reflects The region is marked with the spatial location where the signal strength decays rapidly or the wavefront changes abruptly.

[0020] The goal of the first stage of network model training is to minimize the predicted graph. With timing singularity region map The error between them can be described by the following optimization problem: ; in, The loss function, such as mean squared error (MSE), is used to accurately reconstruct the details of the electromagnetic singularity distribution map.

[0021] After training, the output is the predicted singular region map. This will serve as a conditional input to the second-stage diffusion model, guiding the generation of the final radio map and enabling high-fidelity modeling of areas with high-frequency signal variations.

[0022] In this embodiment, the trained second-stage network model is obtained by training the initial diffusion model using the second sample as training data and the real radio map as labels, including: Multiple second samples are obtained to construct a training dataset; where the second sample is a radio map spliced ​​together with the first sample and its corresponding output data, and the real radio map corresponding to the first sample is used as the real label; Input a portion of the samples from the training dataset into the first... The second-stage network model to be trained is then trained to obtain the first... The prediction results output during this training process; where the prediction results include the predicted drift term. Predicted noise term The latent representation of reconstruction and output data ; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the second samples of the second-stage network model to be trained are used to calculate the loss and serve as the first... The loss during the training process; among which, the loss of the first training process; The loss in this training process includes the loss of the variational autoencoder decoder. Loss of the diffusion model The loss in the diffusion model includes the loss of the drift term. Loss of noise term and the loss of reconstructing the latent representation ; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the second-stage network model to be trained are obtained. The second-stage network model is to be trained; this process is repeated until the number of training iterations or the degree of convergence meets the preset conditions, at which point the trained second-stage network model is obtained.

[0023] Understandably, the input for training the second-stage network model includes the electromagnetic singularity region contour map output from the first stage. Environmental information Building distribution and base station location Simultaneously, the latent representation generated by the VAE encoder is used in the latent space. The output is a complete radio map. It is used to represent the power distribution or path loss information within a region, thereby enabling the reconstruction of spatial details and signal strength variations.

[0024] The second-stage model employs a conditional diffusion network architecture, similar in design to the first stage. However, it guides the generation of key spatial features by injecting the singular region map output from the first stage as a structural prior into the network. Specifically, the singular region map is input into the network along with environmental and base station information through a cross-attention mechanism to regulate the generation process, ensuring that the model focuses on electromagnetic singular regions and signal abrupt change areas.

[0025] During training, the model's goal is to minimize the predicted radio map. With real radio maps The error between them can be formally expressed as: ; in, This represents the second-stage conditional diffusion model. Indicates trainable parameters, This represents an appropriate loss function, such as mean squared error (MSE), used to accurately reconstruct high-frequency spatial details and signal abrupt change regions.

[0026] In this embodiment, for the radio map, an electromagnetic singularity contour map is constructed, including: Based on the characteristics of electromagnetic wave propagation in a passive, isotropic, and time-invariant linear medium, its behavior can be derived from Maxwell's equations, thus constructing a radio wave propagation model; when considering the steady-state sinusoidal time-dependent form... At this time, the electric field and the magnetic field satisfy the following curl relationship: ; ; in, Represents electric field, Indicates magnetic field, Represents the imaginary unit. Represents angular frequency. Indicates permeability, Indicates the dielectric constant; The electric field satisfies the vector Helmholtz equation, expressed as: ; Within a locally isotropic free space region, the scalar field Satisfying the radial Helmholtz equation, expressed as: ; The spherical wave solution of this equation is: , Indicates amplitude, Indicates wave number; In complex propagation environments, such as those with obstacles, multipath propagation, or diffraction, the equivalent wavenumber in some regions may be imaginary. , At this point, the electromagnetic wave exhibits ephemeris behavior, and its amplitude and power decrease with distance as follows: ; ; To describe the local variation characteristics of electromagnetic fields, the local effective wavenumber index is introduced. , is represented as: ; This indicates that the region exhibits wavefront bending, local energy attenuation, or abrupt changes, i.e., potential electromagnetic singularities or signal abrupt change regions. Represents a spatial location vector; The local effective wavenumber index Binarization is performed to obtain the electromagnetic singularity contour map. , is represented as: ; in, Indicates the row index of the grid. The column index representing the grid.

[0027] Electromagnetic singularity profile It can serve as an intermediate structure for guiding physical information, guiding the network to focus on electromagnetic singularities during the training of diffusion models, thereby achieving high-fidelity reconstruction of high-frequency signal changes and spatial details in radio maps.

[0028] Furthermore, in this embodiment, the VAE is trained, and the VAE's encoder takes the input radio map samples... Mapping to latent representation The decoder then from reconstruction If the training objective is to minimize the lower bound of negative evidence (ELBO), then the loss of the variational autoencoder decoder... The expression is: ; in, Representing reconstruction loss, it measures the difference between the input and the reconstructed output, and is typically expressed using... Norm, This represents a regularization constraint that ensures the latent distribution conforms to the standard normal distribution. Indicates to Seeking expectations, This represents the input data to the variational autoencoder. Representing potential representations, These represent the parameters of the variational autodecoder. The parameters represent the variational autoencoder. Represents a probability distribution. Indicates the first sample or the second sample. This represents the distribution of the variational autodecoder. This represents the prior distribution of the latent variable.

[0029] Furthermore, in this embodiment, the loss of the diffusion model... The expression is: ; in, This represents the loss due to the drift term. This represents the loss due to noise. The loss represents the cost of reconstructing the underlying representation. , and These represent adjustable weight coefficients used to balance the impact of each component on training. By minimizing the total loss, the model can simultaneously optimize the deterministic trajectory and random generation capability of the latent representation, achieving high-fidelity radio map generation.

[0030] Furthermore, in this embodiment, a conditional denoising diffusion model is used for training in the latent space. This model needs to predict two key quantities during the reverse time diffusion process: the drift term... and noise terms The drift term describes the deterministic trajectory of the latent representation, while the noise term compensates for the random perturbations introduced during the forward diffusion stage.

[0031] Loss of drift term The expression is: ; in, Indicates a potential representation.

[0032] Loss of noise term The expression is: ; in, Indicates in Superimposed noise; To enhance the global consistency of the latent representation, this embodiment introduces an auxiliary reconstruction loss, which is achieved by reconstructing the denoised latent variables. With true potential representation In comparison, the mean squared error is calculated, and the loss for reconstructing the latent representation is determined. The expression is: ; in, This represents the potential representation of reconstruction.

[0033] Furthermore, this invention employs a continuous-time decoupled diffusion framework, enabling the model to possess adaptive step-size inference capabilities. This is achieved by adjusting the denoising step size. The model can dynamically control the fineness of the denoised trajectories, thus achieving a balance between generation accuracy and computational efficiency. In real-time applications, fewer steps can be selected for fast inference, or more steps can be used to obtain a high-fidelity radio map. This method ensures... It is both efficient and accurate in dynamic wireless communication environments, and can recover high-frequency spatial details and signal abrupt change regions.

[0034] In summary, this invention provides a radio map reconstruction method based on a dual-diffusion model, employing a two-stage generation architecture. The first stage utilizes environmental information and base station layout to predict the contours of electromagnetic singularities, guiding the first-stage network model to focus on key spatially changing areas. The second stage combines contour information and environmental data to generate a high-fidelity radio map, achieving refined reconstruction of complex spatial features. The proposed method maintains high accuracy in dynamic and changing communication environments while maintaining computational efficiency, meeting the needs of real-time or near-real-time applications. It also demonstrates good generalization ability across different types of radio map scenarios, accurately reconstructing the main propagation path and effectively capturing multipath signals and local signal abrupt changes. This provides reliable environmental perception support for applications such as UAV trajectory planning, intelligent reflector configuration, mobile network optimization, and low-overhead channel estimation. By organically combining physical laws with data-driven generation methods, this invention achieves efficient, accurate, and dynamically adaptable radio map construction, providing a solid technical guarantee for environmental perception, optimization, and intelligent decision-making in future 6G and higher-generation communication networks.

[0035] In an optional embodiment of the present invention, the effectiveness of the radio map reconstruction method based on the dual-diffusion model provided in the above embodiment is verified by simulation experiments, specifically as follows: Experiments were conducted using three radio map datasets with different environmental complexities and electromagnetic propagation characteristics. The datasets were designed to evaluate the effectiveness of radio map reconstruction in static and dynamic radio scenarios, and their base datasets were selected from... The dataset contains 700 independently constructed city maps, each including several buildings. The building data is sourced from... Different maps cover urban, town, and rural environments. Each map specifies 80 transmitter locations and provides corresponding ground truth (GT) path loss values. All maps are converted to 256×256 binary images, where each pixel represents an area of ​​1 square meter, with a pixel value of 1 inside buildings and 0 in open areas. Transmitters and receivers are both 1.5 meters high, buildings are uniformly modeled as 25 meters high, the transmit power is fixed at 23 dBm, and the carrier frequency is 5.9 GHz.

[0036] Specifically, this simulation experiment considers the following three types of radio map variants: 1. Static radio map (SRM), generated based on the Dominant Path Model (DPM), only considers the influence of the main propagation path and large-scale static buildings; 2. Dynamic Radio Map (DRM), also based on DPM, but randomly adds small-scale dynamic obstacles (such as vehicles) to the road to simulate the time-varying dynamic characteristics of the urban environment; 3. Multipath-aware radio map (MRM), generated using Intelligent Ray Tracing (IRT), includes multipath propagation with up to four environmental interactions. Due to computational resource limitations, this dataset only considers static buildings and does not include dynamic obstacles.

[0037] The training set includes 600 different environments, each containing 80 base station locations; the evaluation set uses 100 unique unknown environments, each also containing 80 base station locations, thus forming a cross-environment zero-shot generalization test to measure the model's transferability under new layouts and different building densities.

[0038] To comprehensively evaluate the performance of radio map reconstruction in this simulation experiment, an evaluation system combining classical error metrics and perceptual quality metrics was adopted. Normalized mean square error (NMSE) and root mean square error (RMSE) are used to measure the overall accuracy of the prediction results; structural similarity index (SSIM) is used to evaluate the fidelity of the reconstructed image in terms of local brightness, contrast, and structure, reflecting the spatial characteristics of signal variations; peak signal-to-noise ratio (PSNR) is used to evaluate the reconstruction quality of the image, especially the fidelity of edges and textures. This evaluation system can simultaneously reflect the quantization accuracy and perceptual quality of the reconstruction, providing a comprehensive performance evaluation for radio map generation methods.

[0039] This invention will It was compared with four representative deep learning methods, covering different architectures and features. Based on the U-Net architecture, maps are inferred directly from environmental information through supervised learning; based on generative adversarial networks... ; By combining VAE with denoising U-Net, inverse electromagnetic propagation is modeled in the latent space to achieve high-precision reconstruction; Based on conditional diffusion, environmental geometry and base station information are fused, and physical consistency high-resolution generation is achieved through Helmholtz PDE residuals. All methods are trained and evaluated on the same dataset and under the same experimental conditions.

[0040] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating the comparison of various indicators provided in an embodiment of the present invention. The method demonstrates significant performance improvements across different types of radio maps (SRM, DRM, MRM). Comparative experimental results show that... It outperforms existing baseline models on all evaluation metrics.

[0041] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the comparison of the SRM radio map reconstruction task provided in the embodiments of the present invention. In the SRM scenario, the normalized mean square error (NMSE) of the model is reduced by more than 40% compared with the traditional GAN ​​framework, and the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) reach 0.9773 and 34.46dB, respectively, which fully verifies its excellent ability in preserving signal boundaries and structural consistency.

[0042] Please see Figure 6 , Figure 6 This is a schematic diagram illustrating a comparison of DRM radio map reconstruction tasks provided in an embodiment of the present invention. In more complex DRM scenarios, It also maintained a stable lead. Its NMSE decreased by about 40%, and its RMSE, SSIM, PSNR and other indicators continued to outperform other methods. Especially in environments with dynamic obstacles and multipath scattering effects, the model can still maintain high-fidelity reconstruction performance, demonstrating good generalization and robustness.

[0043] In addition, please see Figure 7 , Figure 7 This is a schematic diagram illustrating a comparison of MRM radio map reconstruction tasks provided in this embodiment of the invention. In the MRM scenario, this method still demonstrates significant advantages. Its NMSE reaches 0.0066, a further reduction of approximately 45% compared to the optimal baseline; RMSE is only 0.0236, and SSIM and PSNR are improved to 0.9674 and 32.68 dB, respectively. The generated radio map has clear boundaries and can accurately capture local high-frequency changes in the electromagnetic field caused by scattering or singularities, greatly improving its applicability in high-precision tasks such as UAV trajectory planning, beamforming, and channel prediction.

[0044] For location performance verification, please refer to Figure 8 , Figure 8 This is a schematic diagram comparing the positioning errors of different methods provided in this embodiment of the invention. To evaluate the practical application value of the radio map generated by this invention, a positioning experiment was conducted using the K-Nearest Neighbor (KNN, K=5) algorithm. 3,000 test points were randomly selected for each map, covering both line-of-sight (LoS) and non-line-of-sight (NLoS) environments. Experimental results show that... It achieves the lowest average positioning error in SRM, DRM, and MRM scenarios, all below 5 meters, which is significantly better than... , and This indicates that the method of the present invention not only improves the quality of radio map reconstruction under complex propagation conditions, but also enhances the downstream positioning accuracy.

[0045] For a comparison of inference efficiency, please refer to [link / reference]. Figure 9 , Figure 9 This is a schematic diagram comparing the inference time and memory usage of different models provided in this embodiment of the invention. Inference latency and memory consumption were tested on an NVIDIA RTX Pro6000 GPU. The results show that traditional shallow models, such as... , Inference time is extremely short, and memory usage is low; while diffusion-based models ( , , Memory usage is about 3 to 4 times higher, but due to LDM architecture optimization, Inference time is still less than 1 second, much faster than RMDM's 19 seconds, making it suitable for near real-time deployment.

[0046] For ablation analysis, please refer to [link / reference]. Figures 10-12 , Figure 10 This invention provides an embodiment that uses different contour maps as... A diagram illustrating the performance comparison of inputs. Figure 11 This is based on the embodiments of the present invention. A schematic diagram of quantitative comparison on a map. Figure 12 This is provided by the embodiments of the present invention. This is a schematic diagram illustrating the qualitative comparison of maps. To verify the importance of physical information guidance, this invention replaces the map contour derived from the Helmholtz equation with traditional edge detection methods (such as Canny and LBP). Experimental results show that the alternative solution exhibits significantly lower performance in metrics such as NMSE, SSIM, and PSNR, even falling below that of the solution without contour guidance. This indicates that, based on physics The index can effectively capture the structure of electromagnetic singularities, providing more reliable conditional information for radio map reconstruction. Furthermore, the standard U-Net and diffusion model are used... Comparison of graph predictions shows that the diffusion model significantly outperforms U-Net in all metrics. Specifically, DTC (Distance-Tolerant Coverage) measures the ability of the predicted map to cover real singularities within a certain allowable spatial error range; DTIoU (Distance-Tolerant Intersection over Union) examines the degree of overlap between the predicted region and the real singularity region; and LPIPS (Learned Perceptual Image Patch Similarity) reflects the similarity in perceptual quality between the predicted and real images, emphasizing texture and detail fidelity. Experimental results show that the diffusion model generates… The images are clearer and more detailed, further validating its advantages in capturing high-frequency electromagnetic exotic structures.

[0047] I. Simulation Conditions The simulation experiment conditions and hardware configuration for this embodiment are as follows: The experiment was conducted on a server equipped with an NVIDIA RTX Pro 6000 GPU, running the Ubuntu operating system. Dataset: The publicly available benchmark dataset RadioMapSeer was used, and each scene included a building layout matrix. This indicates the location of static obstacles. (Base station location) The corresponding true path loss matrix is ​​also included. Dynamic obstacles (such as vehicles) are simulated using randomly generated reflection coefficients to represent multipath effects. Data split: 600 scenarios are used for training, and 100 scenarios are used for testing.

[0048] II. Simulation Content and Result Analysis Traditional methods: full-wave electromagnetic simulation calculation (based on numerical solution of Maxwell's equations); data-driven methods: RadioUNet (CNN architecture), RME-GAN (Generative Adversarial Network), RadioDiff (diffusion model), RMDM (physical diffusion model).

[0049] Evaluation metrics: Accuracy metrics: NMSE (Normalized Mean Square Error), RMSE (Root Mean Square Error), PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity), DTC (Distance Tolerance Coverage), DTIoU (Distance Tolerance Cross-Union Ratio), LPIPS (Perceptual Patch Similarity). Efficiency metrics: Sub-inference time (seconds).

[0050] Example 1: Verification of singularity extraction guided by physical information.

[0051] Objective: Verification middle The index demonstrates the ability to locate electromagnetic singularities (regions of abrupt changes in high gradient / path loss). Implementation method: Calculate the true RM using the Helmholtz equation. Values ​​are used to generate a binary singularity map; a diffusion model is trained to predict... The figure shows the loss function as MSE.

[0052] Results (e.g.) Figure 11 (As shown): On the SRM, DRM, and MRM datasets, the diffusion model predicts... The graph significantly outperforms the standard U-Net across all metrics. Specifically, the DTC reaches 0.9703, 0.9782, and 0.9577, while U-Net's corresponding values ​​are only 0.8141, 0.6857, and 0.7433; the DTIoU is 0.9023, 0.8403, and 0.8615, also higher than U-Net. The LPIPS are 0.0933, 0.1107, and 0.1313, significantly lower than U-Net's 0.2042, 0.2449, and 0.2590. Visualization results show that the diffusion model generates... The image boundaries are clearer and sharper, with richer details, enabling accurate location of electromagnetic singularity regions and effective capture of high-frequency abrupt structural changes.

[0053] Example 2: Radio map reconstruction performance.

[0054] Scenario 1: Static Dominant Path (DPM) Results (e.g.) Figure 4 and Figure 5 As shown): The NMSE is 0.0043, which is relatively... The 0.0072 was reduced by approximately 40%; PSNR=34.46dB, SSIM=0.9773; RM edges and details were clearly visible.

[0055] Scenario 2: Dynamic / Multi-path Complex Environment (DRM / MRM) Results (e.g.) Figure 4 , Figure 5 , Figure 6 As shown in the figure: In the DRM dataset, the reconstruction results have an NMSE of 0.0054, an RMSE of 0.0208, an SSIM of 0.9704, and a PSNR of 33.79; in the MRM(IRT) dataset, the NMSE is 0.0066, the RMSE is 0.0236, the SSIM is 0.9674, and the PSNR is 32.68. It is evident that the method of this invention, even in occluded areas or environments with strong multipath interference, can maintain the fine structure and signal abrupt change characteristics of the radio map during reconstruction, and its overall performance is significantly better than... , and Existing methods, etc.

[0056] Example 3: Real-time verification.

[0057] Test content (e.g.) Figure 9 (As shown): Single inference time of different methods on a 256×256 grid. Results (as shown) Figure 4 As shown): In terms of reasoning efficiency, Its single inference time is 0.76 seconds, with a memory usage of 3.02GB, a significant improvement over RMDM's 19.3 seconds; although slightly higher than... It takes 0.24 seconds, but it is still sufficient to meet the needs of near real-time deployment.

[0058] Example 4: Positioning performance evaluation.

[0059] Test content: Conduct localization experiments on the SRM, DRM, and MRM datasets, randomly selecting each map. Figure 3 ,000 test points, including LoS (line-of-sight) and NLoS (non-line-of-sight) locations. Localization was performed using the KNN algorithm (K=5) based on the generated radio map, and the model's localization accuracy was evaluated under dynamic multipath and occlusion conditions.

[0060] Test results (e.g.) Figure 8 As shown): The average positioning errors on SRM, DRM, and MRM were 3.32m, 3.87m, and 4.72m, respectively, which are significantly lower than those on SRM, DRM, and MRM. , and Equal comparison method ( (5.90m, 7.18m, 8.87m). Experiments show that even in complex environments, physical information-guided... The graph structure can preserve the fine features of radio map reconstruction, thereby ensuring high accuracy for downstream positioning tasks.

[0061] Based on the same inventive concept, the present invention also provides a radio map reconstruction apparatus based on a dual-diffusion model, used to implement the radio map reconstruction method based on a dual-diffusion model provided in the above embodiments of the present invention. Embodiments of the method can be referred to above and will not be repeated here. The apparatus includes: The data acquisition module is used to acquire data to be processed; the data to be processed includes environmental information, building distribution, and base station locations. The data processing module is used to process the data to be processed using a pre-trained first-stage network model, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities; and to process the electromagnetic singularities and the data to be processed using a pre-trained second-stage network model, generating a radio map using the electromagnetic singularities as a structural guide. The first-stage network model, trained using the first sample as training data and a pre-constructed electromagnetic singularity contour map as labels, is used to train the initial diffusion model. The second-stage network model, trained using the second sample as training data and a real radio map as labels, is used to train the initial diffusion model.

[0062] In summary, the radio map reconstruction method and apparatus based on a dual-diffusion model provided by this invention have the following beneficial effects, including: This invention proposes a radio map construction method based on the Helmholtz equation and a dual-network architecture. It has the following significant advantages in 6G environment-aware communication: First, high-precision radio map reconstruction.

[0063] The present invention proposes This method effectively captures the structure of electromagnetic singularities through a physically-guided diffusion model, enabling high-precision radio map reconstruction for various scenarios such as SRM, DRM, and MRM. Experimental results show that in complex multipath environments, the method outperforms existing methods in terms of NMSE, RMSE, SSIM, and PSNR. For example, in DRM, the NMSE is only 0.0054 and the SSIM reaches 0.9704; in MRM (IRT), the NMSE is 0.0066 and the PSNR is 32.68 dB, maintaining fine structure and texture details in occluded areas or areas of signal abrupt changes.

[0064] Second, the advantages of physical information guidance.

[0065] Derived from the Helmholtz equation This invention provides a map that accurately locates EM singularities and offers reliable conditional information. Compared to traditional edge detection methods (such as Canny and LBP) or contour-free guided methods, this invention improves upon NMSE, SSIM, PSNR, and other performance metrics. It outperforms in all graph prediction metrics (DTC, DTIoU, LPIPS) and generates better results. Clear map boundaries and rich details help enhance the physical consistency of radio map reconstruction.

[0066] Third, dynamic environment adaptability and real-time performance.

[0067] The method of this invention also exhibits stable performance in dynamic multipath environments. Through pre-trained models and real-time training deployment, it can quickly respond to changes in path loss caused by moving obstacles, achieving real-time updates of the radio map. On a 256×256 grid, the single inference time is 0.76 seconds, with a GPU memory usage of 3.02GB, which is a significant improvement compared to the Physically Guided Diffusion Model (RMDM) (19.3 seconds), while meeting the requirements of near real-time applications.

[0068] Fourth, downstream positioning performance is improved.

[0069] The high-precision radio maps generated by this invention can be directly used for positioning tasks. Experiments show that, on SRM, DRM, and MRM datasets, based on... The average positioning errors of the KNN were 3.32m, 3.87m, and 4.72m, respectively, which were significantly lower than those of the previous models. , and This enables high-precision positioning in complex environments.

[0070] In summary, this invention achieves high-precision reconstruction and detail preservation of radio maps by introducing a physically-guided diffusion model. It effectively captures electromagnetic singularity structures and maintains stable performance in complex and dynamic wireless environments. This method not only improves the reliability and usability of map reconstruction but also significantly enhances map-based positioning and navigation capabilities, providing strong technical support for UAV path planning, intelligent transportation, and future wireless communication systems.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0072] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0073] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radio map reconstruction method based on a dual-diffusion model, characterized in that, include: Acquire data to be processed; wherein, the data to be processed includes environmental information, building distribution, and base station locations; A pre-trained first-stage network model is used to process the data to be processed, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities. A pre-trained second-stage network model is then used to process both the electromagnetic singularities and the data to be processed, using the electromagnetic singularities as a structural guide to generate a radio map. The first-stage network model, which has been trained, is obtained by training the initial diffusion model using the first sample as training data and the pre-constructed electromagnetic singularity contour map as labels. The second-stage network model, which has been trained, is obtained by training the initial diffusion model using the second sample as training data and the real radio map as labels.

2. The radio map reconstruction method based on the dual-diffusion model according to claim 1, characterized in that, The architecture of the trained first-stage network model is the same as that of the trained second-stage network model, both including: a variational autoencoder, a diffusion model, and a variational autodecoder; wherein the diffusion model is a conditional denoising diffusion model. The variational autoencoder is used to process the input data. Mapping to a lower-dimensional space yields the latent representation. And add noise to the potential representation. ; The diffusion model is used to represent the potential of superimposed noise. The process involves modeling in the latent space, preserving the structural features of electromagnetic singularities and high-curvature regions through latent representation, and obtaining the predicted drift term. and predicted noise term Based on the predicted drift term and the predicted noise term, the reconstructed latent representation is obtained. , is represented as: ； The variational autodecoder is used to reconstruct the reconstructed latent representation to obtain output data. .

3. The radio map reconstruction method based on the dual-diffusion model according to claim 1, characterized in that, The trained first-stage network model is obtained by training the initial diffusion model using the first sample as training data and a pre-constructed electromagnetic singularity contour map as labels, including: Multiple first samples are obtained to construct a training dataset; wherein, the first sample is a radio map including environmental information, building distribution and base station location; For radio maps, an electromagnetic singularity contour map is constructed as the true label for the first sample; Input a portion of the samples from the training dataset into the first... The first-stage network model to be trained is then trained to obtain the second... The prediction results output during the training process; wherein, the prediction results include the predicted drift term. Predicted noise term The latent representation of reconstruction and output data ; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the first samples of the first-stage network model to be trained are used to calculate the loss and serve as the basis for the second-stage training. The loss during the training process; among which, the loss of the first training process; The loss in this training process includes the loss of the variational autoencoder decoder. Loss of the diffusion model The loss of the diffusion model includes the loss of the drift term. Loss of noise term and the loss of reconstructing the latent representation ; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the first-stage network model to be trained are obtained. The first-stage network model to be trained is then iterated until the number of training iterations or the degree of convergence meets the preset conditions, thus obtaining the trained first-stage network model.

4. The radio map reconstruction method based on the dual-diffusion model according to claim 1, characterized in that, The trained second-stage network model is obtained by training the initial diffusion model using the second sample as training data and real radio maps as labels, including: Multiple second samples are obtained to construct a training dataset; wherein, the second sample is a radio map spliced ​​together with the first sample and its corresponding output data, and the real radio map corresponding to the first sample is used as the real label; Input a portion of the samples from the training dataset into the first... The second-stage network model to be trained is then trained to obtain the first... The prediction results output during the training process; wherein, the prediction results include the predicted drift term. Predicted noise term The latent representation of reconstruction and output data ; According to the The prediction results output during the training process are the same as those during the training phase. The true labels of the second samples of the second-stage network model to be trained are used to calculate the loss and serve as the first... The loss during the training process; among which, the loss of the first training process; The loss in this training process includes the loss of the variational autoencoder decoder. Loss of the diffusion model The loss of the diffusion model includes the loss of the drift term. Loss of noise term and the loss of reconstructing the latent representation ; According to the The loss from the training process is backpropagated to update the loss from the training process. The network parameters of the second-stage network model to be trained are obtained. The second-stage network model to be trained is then iterated until the number of training iterations or the degree of convergence meets the preset conditions, thus obtaining the trained second-stage network model.

5. The radio map reconstruction method based on the dual-diffusion model according to claim 3, characterized in that, For radio maps, construct an electromagnetic singularity contour map, including: Based on the characteristics of electromagnetic wave propagation in a passive, isotropic, and time-invariant linear medium, a radio wave propagation model is constructed; when considering the steady-state sinusoidal time-dependent form... At this time, the electric field and the magnetic field satisfy the following curl relationship: ； ； in, Represents electric field, Indicates magnetic field, Represents the imaginary unit. Represents angular frequency. Indicates magnetic permeability, Indicates the dielectric constant; The electric field satisfies the vector Helmholtz equation, expressed as: ； Within a locally isotropic free space region, the scalar field Satisfying the radial Helmholtz equation, expressed as: ； in, , Indicates amplitude, Indicates wave number; To describe the local variation characteristics of electromagnetic fields, the local effective wavenumber index is introduced. , is represented as: ； This indicates that the region exhibits wavefront bending, local energy attenuation, or abrupt changes. Represents a spatial location vector; The local effective wavenumber index Binarization is performed to obtain the electromagnetic singularity contour map. , is represented as: ； in, Indicates the row index of the grid. The column index representing the grid.

6. The radio map reconstruction method based on the dual-diffusion model according to claim 3 or 4, characterized in that, The loss of the variational autoencoder decoder The expression is: ； in, Indicates the reconstruction loss. This represents a regularization constraint. Indicates to Seeking expectations, This represents the input data to the variational autoencoder. Representing potential representations, These represent the parameters of the variational autodecoder. The parameters represent the variational autoencoder. Represents a probability distribution. Indicates the first sample or the second sample. This represents the distribution of the variational autodecoder. This represents the prior distribution of the latent variable.

7. The radio map reconstruction method based on the dual-diffusion model according to claim 3 or 4, characterized in that, The loss of the diffusion model The expression is: ； in, This represents the loss due to the drift term. This represents the loss due to noise. The loss represents the cost of reconstructing the underlying representation. , and These represent adjustable weighting coefficients.

8. The radio map reconstruction method based on the dual-diffusion model according to claim 7, characterized in that, The loss of the drift term The expression is: ； in, Representing a potential representation; The loss of the noise term The expression is: ； in, Indicates in Superimposed noise; The loss of the reconstructed latent representation The expression is: ； in, This represents a potential representation of reconstruction.

9. A radio map reconstruction device based on a dual-diffusion model, characterized in that, include: The data acquisition module is used to acquire data to be processed; wherein, the data to be processed includes environmental information, building distribution, and base station locations; The data processing module is used to process the data to be processed using a pre-trained first-stage network model, capturing the spatial characteristics of rapidly changing signals and predicting electromagnetic singularities; and to process the electromagnetic singularities and the data to be processed using a pre-trained second-stage network model, using the electromagnetic singularities as structural guidance to generate a radio map; wherein... The first-stage network model, which has been trained, is obtained by training the initial diffusion model using the first sample as training data and the pre-constructed electromagnetic singularity contour map as labels. The second-stage network model, which has been trained, is obtained by training the initial diffusion model using the second sample as training data and the real radio map as labels.

Images