A radio map fast generation method based on flow matching

Abstract

The application discloses a radio map fast generation method based on flow matching, solves the problem of how to improve radio map generation efficiency, and belongs to the field of computer vision; comprising: discretizing a radio map scene into a grid, when dynamic obstacles exist, constructing a first condition matrix through element-by-element maximum value operation; when no dynamic obstacles exist, constructing a second condition matrix through element-by-element maximum value operation; only adding noise to unknown areas in the target condition matrix, constructing a forward process, and obtaining a vector field; introducing a reverse block into a neural network with a U-Net structure to gradually reconstruct a complete initial radio map, constructing a reverse process; projecting the first condition guide tensor or the second condition guide tensor in the initial radio map into an embedding representation through an extractor, injecting the embedding representation into an intermediate layer of the U-Net through a multi-head attention mechanism, generating a radio map, and obtaining RRFlow; and the application improves the radio map generation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of rapid radio map generation technology, and relates to a rapid radio map generation method based on flow matching. Background Technology

[0002] A radio map is a map that visualizes the automated and optimized propagation characteristics of radio signals within a specific area (such as an urban environment) to support 6G network applications such as positioning, optimization, and communication. Figure 1 The image shows a simulation diagram of path loss. Figure 1 The red markers indicate the transmitter's location, and the blue markers indicate the receiver's location. The lines connecting them show the different propagation paths between the transmitter and receiver, including direct, reflected, and diffracted components. These paths collectively constitute a multipath propagation scenario that affects the overall path loss characteristics. Currently, there are two methods for generating radio maps:

[0003] 1. Radio map construction methods based on traditional interpolation and propagation models; for example, the direct interpolation technique Kriging is widely used because it is computationally efficient and does not require prior geographical knowledge; however, Kriging's key statistical assumption of stationary path loss often does not hold in complex urban environments, leading to performance degradation in areas with significant shadows and multipath effects.

[0004] 2. Indirect Methods: This method first estimates transmitter parameters (such as location and power), then uses propagation models such as ray tracing or log-normal shadowing to calculate the received power at unknown points. This method can utilize geographic data (such as building layouts), but its accuracy is highly dependent on the calibration of the propagation model; incorrect model parameters can lead to coverage prediction errors. Deep learning-based radio map construction methods can also be used. For example, RadioUNet, a convolutional neural network based on the U-Net architecture, treats radio map construction as an image-to-image translation problem. It learns environmental features and propagation characteristics through an encoder-decoder structure and uses supervised learning to achieve end-to-end path loss prediction. It is computationally efficient and has fast inference speed, performing well in uniform sampling and static environments, but struggles with detail recovery in non-uniform sampling and dynamic environments. Another method is RadioDiff, which introduces a denoising diffusion probability model (DDPM) to handle the conditional generation task of radio maps. RadioDiff uses a decoupled diffusion mechanism (DDM) and an adaptive Fourier transform (AFT) module to improve training and inference efficiency, and compresses the data dimensionality in the latent space using a variational autoencoder (VAE) to improve computational efficiency. This method demonstrates state-of-the-art performance in both static and dynamic radio map construction, but the inherent randomness of the diffusion model may result in an overall error slightly above the optimal level.

[0005] Traditional radio mapping methods, such as Kriging-based direct interpolation, are computationally efficient and do not rely on prior geographic information. However, their core assumption—the stationarity of path loss statistics—is frequently broken in complex environments such as cities. This leads to a significant drop in predictive performance in areas with pronounced shading and multipath effects. Indirect methods relying on ray tracing or log-normal shading models, while utilizing geographic data, are highly dependent on the fine calibration of the propagation model. Inappropriate model parameter settings can directly cause errors in coverage prediction.

[0006] In recent years, emerging machine learning and deep learning methods, especially models employing discriminative learning paradigms such as RadioUNet, have made progress in handling static environments and extracting features using convolutional neural networks. However, they are relatively weak in handling dynamic environments (such as moving vehicles) and lack effective mechanisms to adapt to changes in signal propagation in these scenarios. Furthermore, discriminative models based on pixel-level loss functions (such as MSE) tend to produce blurry predictions when generating radio maps, especially in critical areas such as obstacle edges, where significant detail loss occurs. Simultaneously, these models also exhibit limitations when dealing with sparse and irregular measurement data common in real-world applications, struggling to accurately reconstruct complete radio maps from a limited number of sampling points.

[0007] Regarding generative models, while Generative Adversarial Networks (GANs) are capable of generating data, their training process is highly unstable and prone to mode collapse. Diffusion models, such as RadioDiff, offer improvements in generation quality and stability, and enhance detail extraction by incorporating modules like Adaptive Fourier Transform (AFT). However, their inherent randomness and iterative sampling processes result in significant computational resource requirements and slow generation speeds, limiting their practicality in applications demanding rapid response or real-time processing.

[0008] For example, in the application number 202510197292.0, the invention title is: "An electromagnetic spectrum map construction method for scenes with undetectable regions". In this method, a partial convolutional U-Net is used to treat spectrum map construction as image inpainting, but it cannot be accurately constructed in extreme environments without sampling conditions.

[0009] For example, the application number is 202410446883.2, and the invention title is: A method for mapping the spectrogram of three-dimensional complex urban environments. The proposed method is a three-dimensional spectrogram mapping method based on UAV optimized sampling and compressed sensing. Its core relies on on-site, active physical sampling data. The entire technical process, from calculating the optimal sampling position to signal recovery, is based on the premise of "sampling". This leads to high data acquisition costs, time overhead and dependence on hardware devices (such as UAVs). Moreover, the method will completely fail in extreme scenarios with no sampling.

[0010] In summary, current methods for constructing radio maps, whether traditional interpolation techniques or discriminative models based on deep learning, face challenges. For example, traditional methods are computationally inefficient or heavily reliant on expensive and often sparse measurement data; while deep learning models, such as RadioUNet, perform reasonably well in static environments but struggle to effectively capture the complexity of dynamic environments (e.g., signal propagation changes caused by moving vehicles). Many discriminative deep learning models based on pixel-level loss functions often produce blurry predictions and lost details when generating radio maps, especially at obstacle edges, which contradicts applications requiring high-precision signal coverage information. Although generative models like GANs and diffusion models such as RadioDiff offer improvements in quality and detail, the former suffers from unstable training, while the latter is computationally expensive and slow, limiting their feasibility in practical deployment. In conclusion, existing technologies exhibit significant performance bottlenecks in handling dynamism, sparsity, detail restoration, and efficiency. Summary of the Invention

[0011] To address the challenge of accurately modeling dynamic radio environments and generating high-resolution, detailed radio maps while significantly reducing computational costs and dependence on measurement data, this invention proposes a rapid radio map generation method based on flow matching. This method constructs a generative flow model, RRFlow, through a flow matching framework, conditional matrix construction, local noise addition, reverse information injection, and a conditional attention mechanism. RRFlow leverages the probabilistic path learning capability of flow matching, combined with a conditional generation mechanism and a local noise strategy, to achieve efficient and high-fidelity radio map reconstruction. This invention provides an efficient, accurate, and robust dynamic radio map generation method that efficiently combines static inputs such as geographic and building information with dynamic obstacle information to generate accurate and finely structured radio maps. This meets the high demands of future communication technologies such as 6G for radio maps and overcomes the shortcomings of existing methods in terms of dynamic environment adaptability, sparse sampling reconstruction quality, and generation efficiency. This method is not only applicable to static environments but also effectively handles path loss variations caused by dynamic obstacles. It can still generate high-quality reconstruction results under extremely sparse sampling or even no sampling conditions, providing a reliable radio map construction solution for applications such as 6G network planning, indoor and outdoor positioning, and intelligent transportation systems.

[0012] The objective of this invention is specifically achieved through the following technical solutions:

[0013] This invention discloses a method for rapid generation of radio maps based on flow matching, the method comprising:

[0014] Step 1: Discretize the radio map scene into a grid. When dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, dynamic obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a first conditional guidance tensor. Integrate the first conditional guidance tensor by taking the maximum value of each element to construct a first conditional matrix. When no dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a second conditional guidance tensor. Integrate the second conditional guidance tensor by taking the maximum value of each element to construct a second conditional matrix. Use either the first conditional matrix or the second conditional matrix as the target conditional matrix.

[0015] Step 2: Add noise only to the unknown regions in the target condition matrix, while keeping the known obstacle regions unchanged, construct the forward process, and obtain the vector field corresponding to the forward process;

[0016] Step 3: Using a U-Net structured neural network to start from the noise distribution, introduce reverse block learning and predict the vector field, and gradually reconstruct the complete initial radio map, thus constructing the reverse process;

[0017] Step 4: Project the first or second conditional guidance tensor in the initial radio map into an embedded representation through the extractor, and inject it into the intermediate layer of U-Net via a multi-head attention mechanism to generate the radio map; thus obtaining the generative flow model RRFlow.

[0018] Step 5: Train RRFlow by minimizing the flow matching loss function and using the AdamW optimizer to obtain the trained RRFlow.

[0019] In step one, the first condition matrix in the target condition matrix M Or the second condition matrix for:

[0020] ;

[0021] ;

[0022] In the formula, The first condition guides the tensor. ; This serves as the guiding tensor for the second condition. ; Let k be the channel index in the conditional guidance tensor, N be the number of grids, and (i,j) be the position coordinates on the grid.

[0023] In step one, the first conditional guiding tensor is:

[0024] ;

[0025] The second conditional guiding tensor is:

[0026] ;

[0027] In the formula, This is a concatenation function; It is a sparse sampling matrix in grayscale image format. A static obstacle matrix in grayscale image format. A dynamic obstacle matrix in grayscale image format. This is a base station location matrix in grayscale image format.

[0028] In step one, the method for processing grayscale image format is as follows:

[0029] S1, the original sparse sampling matrix pass Normalization mapping to The interval is processed into a sparse sampling matrix in grayscale image format. The processing method is as follows:

[0030] ;

[0031] In the formula, This is the minimum value of the original sparse sampling matrix; This represents the maximum value of the original sparse sampling matrix;

[0032] S2, the original static obstacle matrix Static obstacle matrix processed into grayscale image format using binary encoding ;in, Static obstacle elements , indicates that there is a static obstacle at (i,j); , indicates that there is no static obstacle at (i,j);

[0033] S3, the original dynamic obstacle matrix Dynamic obstacle matrix processed into grayscale image format using binary encoding ;in, Dynamic obstacle elements in , indicating that there is a dynamic obstacle at (i,j); , indicating that there is no dynamic obstacle at (i,j);

[0034] S4, the original base station location matrix Base station location matrix processed into grayscale image format using binary encoding ;in, Base station location elements , indicating that a base station exists at (i,j); , indicating that there is no base station at (i,j).

[0035] In step two, the forward process is constructed as follows:

[0036] ;

[0037] ;

[0038] In the formula, This represents the t-th step noise data obtained by combining the t-th step noise data with the target condition matrix M during the forward process, where t is the time step; the t-th step noise data express With Gaussian noise The noise data obtained at step t by element-wise multiplication; To represent the first step in the forward process The first step of the noise data is obtained by combining the noise data with the target condition matrix M. Step-by-step noisy data, It represents the derivative of the total number of steps in time, and in discrete time it represents the reciprocal of the total number of steps from the target distribution to the initial sample distribution; This represents element-wise multiplication. This represents the unknown region in the objective condition matrix. , represented as in the matrix All elements are 1.

[0039] In step two, the vector field corresponding to the forward process is:

[0040] ;

[0041] In the formula, Given the target distribution during the forward process, the data distribution at step t is... Given the target distribution and the initial sample distribution during the forward pass, the data distribution at step t is... For the target distribution, For the initial sample distribution, Let be the data distribution at step t.

[0042] In step three, the reverse block is:

[0043] ;

[0044] In the formula, For the first step in the reverse process The t-dt-th step noise data is obtained by combining the step noise data with the target condition matrix M; This represents the noise data at step t-dt without incorporating the target condition matrix M during the reverse process.

[0045] In step three, the method for constructing the reverse process is as follows:

[0046] ;

[0047] In the formula, For the first step in the reverse process The t-th step noise data is obtained by combining the initial noise data with the target condition matrix M. Let be the vector field that the U-Net structure neural network learns and predicts at step t.

[0048] In step four, the method for injecting the multi-head attention mechanism into the intermediate layers of U-Net is as follows:

[0049] ;

[0050] In the formula, To extract attention from the first or second condition-guided tensor in the initial radio map using a neural network with a U-Net structure. , and These are the queries, keys, and values ​​obtained by projecting the intermediate feature maps of a U-Net-structured neural network. The standardized exponential function transforms any set of numbers into a probability distribution through exponential operations and normalization; T is the transpose; d is the query term. s and keys The dimension is used to scale the dot product.

[0051] In step five, the flow matching loss function is:

[0052] ;

[0053] In the formula, To match the loss function value to the flow, For the parameters of the U-Net neural network structure, This indicates the expectation of a predefined function; This indicates that, in a known reverse process, t follows a uniform distribution in [0,1]. ; For the forward process given the intermediate hidden variable z obey distributed, Given the latent variable z The distribution Let the hidden variable z follow The distribution, Let z be the distribution of the latent variable. Let z be the data distribution at step t given the latent variable z in the forward process, where z is the latent variable. Connect... and .

[0054] In step five, RRFlow is trained on an NVIDIA GeForce RTX 3090 Ti GPU.

[0055] The beneficial effects of this invention are:

[0056] 1. This invention introduces the Flow Matching framework into RRFlow, the construction method for radio map RM, replacing traditional diffusion models or generative adversarial networks. This achieves a more stable and efficient training process, significantly reducing sampling steps. This allows the model to perform rapid inference without sacrificing generation quality, meeting the stringent real-time requirements of practical applications. This invention constructs RRFlow as a probabilistic path learning problem. Through its powerful generation capabilities, it can directly generate high-precision radio maps based solely on prior environmental information (such as building layout and base station locations) without any on-site sampling. This achieves a significant leap in applicability, efficiency, and cost, especially suitable for practical applications requiring large-scale deployment, rapid response, or where physical sampling is infeasible. It represents a leap from local repair relying on sampled data to global generation based on environmental understanding. This innovation enables the technical solution disclosed in this invention to generate high-precision radio maps based solely on environmental information under extreme conditions of no sampling, completely overcoming the heavy reliance on sampled data in existing technologies.

[0057] 2. To address the unique challenges of dynamic environments, this invention proposes an explicit dynamic obstacle modeling mechanism. By introducing a dynamic obstacle matrix that participates in the condition generation process alongside the static obstacle matrix, the model can accurately capture and reconstruct local time-varying path loss changes caused by moving objects (such as vehicles). This design significantly improves the model's adaptability and robustness in complex real-world scenarios. Through explicit dynamic obstacle modeling and efficient flow matching inference, this invention not only significantly enhances adaptability to dynamic environments but also improves generation speed, laying the foundation for real-time applications.

[0058] 3. This invention designs an innovative conditional generation architecture. This architecture integrates multi-source prior information by constructing a unified target condition matrix and uses a conditional injection method based on a multi-head attention mechanism to dynamically fuse the above-mentioned structured target condition information into the generation process of the U-Net backbone network. This ensures that the generated map is highly consistent with all input conditions in terms of detail, so as to produce reliable reconstruction results even under extreme conditions of extremely low sampling rate or even no sampling.

[0059] 4. Regarding noise scheduling and reconstruction strategies, this invention employs a local noise addition and known information retention mechanism. During the forward process, noise is added only to unknown areas. In the entire reverse denoising process, starting from the noise distribution, a reverse block is introduced to ensure that information in known obstacle areas is re-injected, guaranteeing that key information such as known measurements, obstacle outlines, and base station locations are not destroyed during the generation process. This strategy effectively avoids the edge blurring and detail loss problems common in traditional methods, and in particular improves the reconstruction accuracy of signal attenuation changes at obstacle boundaries.

[0060] 5. The Reverse Block forces the re-injection of known information in each backsampling step, further enhancing the model's ability to utilize contextual information. This component has been shown in ablation experiments to play a crucial role in improving the final reconstruction quality, especially edge sharpness and local consistency.

[0061] 6. From the perspective of overall architecture, this invention successfully combines the theoretical advantages of flow matching with the U-Net architecture in computer vision and the attention mechanism in Transformer, creating a generative model that can accurately learn the data distribution and flexibly respond to complex conditional constraints. This model is not only suitable for generation tasks under complete conditions, but also can handle map completion and estimation problems under partial or no observation conditions very well.

[0062] 7. This invention achieves significant optimization in training strategy. By minimizing the flow matching loss function and combining it with the AdamW optimizer to train RRFlow, RRFlow can be trained on a single high-performance GPU. The time and computational resources required are far less than those of the diffusion-based baseline method, which makes this method more practical and scalable.

[0063] The invention's effects are reflected in multiple dimensions. Its core advantage lies in achieving significant breakthroughs in the accuracy, efficiency, robustness, and usability of radio map (RM) construction through the introduction of a flow matching framework and a series of innovative designs. Attached Figure Description

[0064] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0065] Figure 1 This is a simulation diagram of path loss provided in an embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the model structure provided in an embodiment of the present invention.

[0067] Figure 3This is a schematic diagram of the reverse block in the reverse process provided in the embodiment of the present invention.

[0068] Figure 4 This is a qualitative comparison diagram of constructing a radio map using 10 sampling points under conditions without dynamic obstacles, provided by an embodiment of the present invention.

[0069] Figure 5 This is a qualitative comparison diagram of constructing a radio map using 20 sampling points under conditions without dynamic obstacles, provided by an embodiment of the present invention.

[0070] Figure 6 This is a qualitative comparison diagram of constructing a radio map using 50 sampling points under conditions without dynamic obstacles, provided by an embodiment of the present invention.

[0071] Figure 7 This is a qualitative comparison diagram of constructing a radio map under 10 sampling points under dynamic obstacle conditions provided in an embodiment of the present invention.

[0072] Figure 8 This is a qualitative comparison diagram of constructing a radio map under 20 sampling points under dynamic obstacle conditions provided in an embodiment of the present invention.

[0073] Figure 9 This is a qualitative comparison diagram of constructing a radio map under 50 sampling points under dynamic obstacle conditions provided in an embodiment of the present invention.

[0074] Figure 10 This is a qualitative comparison diagram of constructing a radio map under no sampling conditions and no dynamic obstacles, provided by an embodiment of the present invention.

[0075] Figure 11 This is a qualitative comparison diagram of constructing a radio map under conditions of dynamic obstacles without sampling, provided by an embodiment of the present invention.

[0076] Figure 12 This is a comparative schematic diagram of the construction of radio maps under ablation experiments provided in the embodiments of the present invention. Detailed Implementation

[0077] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0078] like Figure 2As shown, this embodiment of the invention provides a method for rapid generation of radio maps based on flow matching, the method comprising:

[0079] Step 1: Discretize the radio map scene into a grid. When dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, dynamic obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a first conditional guidance tensor. Integrate the first conditional guidance tensor by taking the maximum value of each element to construct a first conditional matrix. When no dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a second conditional guidance tensor. Integrate the second conditional guidance tensor by taking the maximum value of each element to construct a second conditional matrix. Use either the first conditional matrix or the second conditional matrix as the target conditional matrix.

[0080] Step 2: Add noise only to the unknown regions in the target condition matrix, while keeping the known obstacle regions unchanged, construct the forward process, and obtain the vector field corresponding to the forward process;

[0081] Step 3: Using a U-Net structured neural network to start from the noise distribution, introduce reverse block learning and predict the vector field, and gradually reconstruct the complete initial radio map, thus constructing the reverse process;

[0082] Step 4: Project the first or second conditional guidance tensor in the initial radio map into an embedded representation through the extractor, and inject it into the intermediate layer of U-Net via a multi-head attention mechanism to generate the radio map; thus obtaining the generative flow model RRFlow.

[0083] Step 5: Train RRFlow by minimizing the flow matching loss function and using the AdamW optimizer to obtain the trained RRFlow.

[0084] In step one, the first condition matrix in the target condition matrix M Or the second condition matrix for:

[0085] ;

[0086] ;

[0087] In the formula, The first condition guides the tensor. ; This serves as the guiding tensor for the second condition. ; Let k be the channel index in the conditional guidance tensor, N be the number of grids, and (i,j) be the position coordinates on the grid.

[0088] The target condition matrix is ​​used to identify known obstacle areas, ensuring that the values ​​of these areas are preserved during noise addition and reconstruction.

[0089] In step one, the first conditional guiding tensor is Tensor:

[0090] ;

[0091] The second condition guides the tensor to be... Tensor:

[0092] ;

[0093] In the formula, This is a concatenation function; It is a sparse sampling matrix in grayscale image format. A static obstacle matrix in grayscale image format. A dynamic obstacle matrix in grayscale image format. This is a base station location matrix in grayscale image format.

[0094] The first and second conditional guiding tensors enable the model to utilize all available information for inference, and the gradients draw from multiple information streams, resulting in more stable training. This unified representation facilitates feature extraction and fusion in convolutional neural networks, improving training stability and convergence speed.

[0095] In step one, the method for processing grayscale image format is as follows:

[0096] S1, the original sparse sampling matrix pass Normalization mapping to The interval is processed into a sparse sampling matrix in grayscale image format. The processing method is as follows:

[0097] ;

[0098] In the formula, This is the minimum value of the original sparse sampling matrix; This represents the maximum value of the original sparse sampling matrix;

[0099] S2, the original static obstacle matrix Static obstacle matrix processed into grayscale image format using binary encoding ;in, Static obstacle elements , indicates that there is a static obstacle at (i,j); , indicates that there is no static obstacle at (i,j);

[0100] S3, the original dynamic obstacle matrix Dynamic obstacle matrix processed into grayscale image format using binary encoding ;in, Dynamic obstacle elements in , indicating that there is a dynamic obstacle at (i,j); , indicating that there is no dynamic obstacle at (i,j);

[0101] S4, the original base station location matrix Base station location matrix processed into grayscale image format using binary encoding ;in, Base station location elements , indicating that a base station exists at (i,j); , indicating that there is no base station at (i,j).

[0102] In step two, the method for constructing the forward process is as follows:

[0103] ;

[0104] ;

[0105] In the formula, This represents the t-th step noise data obtained by combining the t-th step noise data with the target condition matrix M during the forward process, where t is the time step; the t-th step noise data express With Gaussian noise The noise data obtained at step t by element-wise multiplication; To represent the first step in the forward process The first step of the noise data is obtained by combining the noise data with the target condition matrix M. Step-by-step noisy data, It represents the derivative of the total number of steps in time, and in discrete time it represents the reciprocal of the total number of steps from the target distribution to the initial sample distribution; This represents element-wise multiplication. This represents the unknown region in the objective condition matrix. , represented as in the matrix All elements are 1. In step two, the vector field corresponding to the forward process is:

[0106] ;

[0107] In the formula, Given the target distribution during the forward process, the data distribution at step t is... Given the target distribution and the initial sample distribution during the forward pass, the data distribution at step t is... For the target distribution, For the initial sample distribution, Let be the data distribution at step t.

[0108] In step three, the reverse block is:

[0109] ;

[0110] In the formula, For the first step in the reverse process The t-dt-th step noise data is obtained by combining the step noise data with the target condition matrix M; This represents the noise data at step t-dt without incorporating the target condition matrix M during the reverse process.

[0111] like Figure 3 As shown, in step three, the method for constructing the reverse process is as follows:

[0112] ;

[0113] In the formula, For the first step in the reverse process The t-th step noise data is obtained by combining the initial noise data with the target condition matrix M. Let t be the vector field, such as the velocity field, that the neural network with the U-Net structure learns and predicts at step t.

[0114] The initial radio map can achieve deterministic or very few steps of sampling, which significantly improves the generation efficiency and overcomes the time-consuming drawback of iterative sampling in the diffusion model.

[0115] In step four, the method for injecting the multi-head attention mechanism into the intermediate layers of U-Net is as follows:

[0116] ;

[0117] In the formula, To extract attention from the first or second condition-guided tensor in the initial radio map using a neural network with a U-Net structure. , and These are the queries, keys, and values ​​obtained by projecting the intermediate feature maps of a U-Net-structured neural network. The standardized exponential function transforms any set of numbers into a probability distribution through exponential operations and normalization; T is the transpose; d is the query term. s and keys The dimension is used to scale the dot product.

[0118] The introduced conditional attention mechanism further enhances the model's adaptability to complex environmental conditions.

[0119] In step five, the flow matching loss function is:

[0120] ;

[0121] In the formula, To match the loss function value to the flow, For the parameters of the U-Net neural network structure, This indicates the expectation of a predefined function; This indicates that, in a known reverse process, t follows a uniform distribution in [0,1]. ; For the forward process given the intermediate hidden variable z obey distributed, Given the latent variable z The distribution Let the hidden variable z follow The distribution The distribution of the latent variable z; Let z be the data distribution at step t given the latent variable z in the forward process, where z is the latent variable. Connect... and .

[0122] In step five, RRFlow is trained on an NVIDIA GeForce RTX 3090 Ti GPU. During training, the AdamW optimizer is used, and the learning rate is set from... Gradually down to The batch size was set to 4, and training took approximately 24 hours on an NVIDIA GeForce RTX 3090 Ti GPU. The input image resolution was uniformly set to 256×256, and the time step T of the reverse process was set to 1000. The preferred loss function was... Loss measures the pixel-level error between the predicted value and the actual path loss value.

[0123] To verify the effects of the technical solution of the present invention, the following verification experiment is provided:

[0124] The verification experiments of this invention illustrate the implementation process and effects of the RRFlow method through detailed experiments under different conditions. Each experiment was conducted on a test map of the RadioMapSeer dataset, corresponding to a physical region of 256m × 256m, represented by a 256 × 256 pixel grid. All experiments used a pre-trained RRFlow model, whose parameters had been optimized on 40,000 training maps and adjusted on 8,000 verification maps. Finally, the trained model was run on 8,000 test maps to obtain the model's NMSE, RMSE, SSIM, and PSNR metrics on the test set. The experimental environment consisted of Python 3.9 and PyTorch 1.12.1, and the hardware platform was an NVIDIA GeForce RTX 3090 Ti GPU.

[0125] Verification Experiment 1: Static Environment Reconstruction with Different Sampling Points and No Dynamic Obstacles:

[0126] This experiment aims to verify the model's performance in typical sparsely sampled static scenarios. The experimental conditions are as follows: the data is collected from the entire test set, where each image contains building layouts. and a base station (The same applies below) On this map, 10, 20, and 50 sampling points are randomly generated, and their actual path loss values ​​are recorded to form a sparse sampling matrix. This example assumes there are no dynamic obstacles, therefore The zero matrix is ​​used to construct the second conditional matrix. During the forward pass, noise is only added to the unknown regions marked as 0 in M. During the backward generation, the model uses pre-trained weights with a learning rate of 1e-4. In the flow steps with a total number of steps T1=1000, the flow field is predicted and reconstructed step by step using U-Net. Quantitative comparison data of the radio map constructed at different sampling points under the condition of no dynamic obstacles are shown in Table 1.

[0127] Table 1

[0128]

[0129] Qualitative results such as Figures 4 to 6 As shown, by Figures 4 to 6 As can be seen, the reconstruction results of the generated electrical map and the real map clearly show the signal attenuation gradient caused by building shadows, and the error between the predicted value and the real value at the sampling point is very small, which proves the high-precision reconstruction capability of the model under static sparse sampling.

[0130] Verification Experiment 2: Dynamic Environment Reconstruction with Different Sampling Points and Dynamic Obstacles:

[0131] This experiment was conducted to test the robustness of the model in a dynamic environment. The experimental conditions were the same as in validation experiment one, but a vehicle was additionally introduced as a dynamic obstacle. Set the corresponding position in the matrix to 1. Construct the first condition matrix, and use the same training and inference parameters as in verification experiment one.

[0132] Table 2 shows the quantitative evaluation of the radio maps constructed at different sampling points under dynamic obstacle conditions:

[0133] Table 2

[0134]

[0135] Qualitative results such as Figures 7 to 9 As shown, qualitative results indicate that the model successfully reconstructs the local path loss variations caused by the vehicle: in the area around the vehicle, this is consistent with the physical phenomenon of partial electromagnetic wave blockage. Although the introduction of dynamic obstacles increases complexity, the performance metrics only decrease slightly and are significantly better than the baseline (e.g., RadioUNet has a PSNR below 35 dB under the same conditions), validating the model's effective adaptation to dynamic disturbances.

[0136] Verification Experiment 3: Generation Task with Dynamic Obstacles but No Sampling Measurements

[0137] This experiment demonstrates the model's generation capability under the most extreme conditions (completely unsampled). The test map is the same as in validation experiment two, including buildings, base stations, and the same dynamic vehicle obstacles. However, no sampling points are provided this time. A zero matrix (here, it is simply a matrix of zero) (As a matrix of all zeros, it is concatenated). The first conditional matrix is ​​constructed, and the model is unconditionally generated based solely on environmental features and base station locations. The reverse process starts from pure noise and injects it through a conditional attention mechanism. , , Information is used to gradually generate a complete map. Quantitative comparative data on the construction of radio maps under different environments without sampling are shown in Table 3:

[0138] Table 3

[0139]

[0140] The generated result is as follows Figure 10 and Figure 11As shown, the generated map maintains a visually reasonable spatial structure: signal shadows are visible behind buildings, local attenuation occurs at vehicle locations, and the overall path loss distribution conforms to the physical laws of free space propagation and obstacle occlusion. Although the quantitative indicators are lower than those with sampling (NMSE 0.0081, RMSE 0.0189, SSIM 0.9318, PSNR 34.69 dB), this demonstrates that the present invention can still produce reliable estimation results even in the absence of measurement data.

[0141] Verification Experiment 4: Ablation Experiment of Reverse Block:

[0142] This experiment aims to verify the crucial role of the Reverse Block (RB) in the RRFlow model proposed in this invention. The core function of the Reverse Block is to enforce the execution of the reverse sampling process at each step. The operation ensures that known information (such as measurements, obstacle outlines, and base station locations) is re-injected, thereby avoiding information loss during the generation process. This experiment was conducted on the same test map from the RadioMapSeer dataset, which contains complex building layouts. A base station And simulated dynamic vehicle obstacles The experimental setup was identical to the "20 sampling points and dynamic obstacles" conditions in Experiment 2 to ensure fairness in the comparison.

[0143] Validation Experiment 4 compared the full RRFlow model with the inverse block to a simplified version with the inverse block removed. In the simplified version, the inverse process directly updates the flow field predicted by U-Net, instead of performing a forced injection of known information. The quantitative results of the ablation experiment of the inverse block in RRFlow are shown in Table 4.

[0144] Table 4

[0145]

[0146] Qualitative results such as Figure 12 As shown, the qualitative results of the ablation experiment clearly demonstrate the key role of the reverse block (RB) mechanism in radio map reconstruction.

[0147] like Figure 7As shown, when the reverse block is removed, the reconstructed radio map exhibits significant blurring and distortion in obstacle edge regions, particularly failing to accurately capture path loss variations around dynamic obstacles. In contrast, the full RRFlow model, by systematically integrating known environmental information (including building layout, dynamic obstacle locations, and base station coordinates) in each reverse step using reverse blocks, significantly improves the edge sharpness and detail fidelity of the reconstruction.

[0148] Based on comprehensive verification tests one through four, the beneficial effects of the verification tests disclosed in this invention are as follows:

[0149] Regarding reconstruction accuracy, this invention was systematically validated on the publicly available large-scale dataset RadioMapSeer. The experimental conditions included: the dataset was divided into 500 training images, 100 validation images, and 100 test images based on geographical regions; the input images had a uniform resolution of 256×256 pixels; and the evaluation metrics included Normalized Mean Square Error (NMSE), Root Mean Square Error (RMSE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR) to comprehensively measure numerical accuracy, structural fidelity, and perceptual quality. Under the most challenging sparse sampling conditions (10, 20, and 50 sampling points), the RRFlow model of this invention achieved optimal or near-optimal performance. Specifically, in a static environment with 50 sampling points and no dynamic obstacles, RRFlow achieved excellent results with NMSE 0.0046, RMSE 0.0106, SSIM 0.9671, and PSNR 37.83 dB. Its RMSE and PSNR values ​​significantly outperformed all comparison baselines, demonstrating that its generated path loss values ​​have higher accuracy in terms of absolute error and signal fidelity. Of particular note is that in complex scenarios with dynamic obstacles, the performance of traditional methods generally degrades significantly, while RRFlow demonstrates excellent robustness, maintaining a high level of RMSE of 0.0135 and PSNR of 37.53 dB even with 20 dynamic obstacle sampling points. This indicates that the present invention has a more accurate ability to capture and reconstruct local path loss changes caused by dynamic environments. Compared with RadioDiff, which is based on a diffusion model, although RadioDiff has a slight advantage in the structural index SSIM due to its Adaptive Fourier Transform (AFT) module, RRFlow, with its deterministic flow matching framework, performs better in terms of overall error RMSE and signal-to-noise ratio (PSNR), and eliminates the instability caused by the inherent randomness of the diffusion model.

[0150] In terms of generation and computational efficiency, this invention demonstrates significant advantages. Experimental results show that RRFlow's sampling process eliminates the need for iterative multi-step denoising, achieving near-one-step rapid generation. Its inference speed is two orders of magnitude faster than the RadioDiff model, which requires 1000 sampling steps, laying the foundation for real-time applications such as online network planning and dynamic navigation. More importantly, training efficiency is fundamentally improved: RRFlow can be trained in approximately 24-48 hours on a single NVIDIA GeForce RTX 3090 Ti GPU, while the comparative RadioDiff model requires approximately 480 hours on four NVIDIA A100 GPUs. This means that while achieving competitive performance, this invention reduces training computational resource consumption by over 90%, significantly lowering the barrier to entry and cost of model development, demonstrating substantial savings in energy consumption and raw materials (computing hardware).

[0151] In terms of operational complexity and practicality, this invention simplifies the cumbersome multi-stage processing flow of traditional methods (such as the two-stage training of RME-GAN) through an end-to-end flow matching framework and a unified condition matrix design. Users only need to provide an environment layout diagram. Dynamic obstacle information ( ), base station location (R) and optional sparse measurements ( The model can automatically output high-precision radio maps, which greatly reduces the complexity of operation and control, reduces the reliance on professional knowledge and manual intervention, and thus effectively reduces labor intensity.

[0152] Finally, the useful performance extensions of this invention extend beyond the pure RM reconstruction task. The high-precision, high-resolution maps it generates provide a more reliable data foundation for downstream applications in 6G networks, such as spectrum resource management, energy consumption optimization (reducing excessive base station deployment through accurate coverage prediction), seamless indoor and outdoor positioning, and intelligent transportation systems (e.g., V2X link stability prediction). This promotes improved communication efficiency and energy conservation at the system level, offering certain environmental benefits. In summary, through theoretical and experimental verification, this invention not only leads in all aspects of technical indicators but also demonstrates significant advantages in efficiency, cost, and ease of use, providing a novel and practical solution for the construction of dynamic radio maps.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for rapid generation of radio maps based on flow matching, characterized in that, The method includes: Step 1: Discretize the radio map scene into a grid. When dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, dynamic obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a first conditional guidance tensor. Integrate the first conditional guidance tensor by taking the maximum value of each element to construct a first conditional matrix. When no dynamic obstacles exist, process the sparse sampling matrix, static obstacle matrix, and base station location matrix in the grid into grayscale image format and then concatenate them into a second conditional guidance tensor. Integrate the second conditional guidance tensor by taking the maximum value of each element to construct a second conditional matrix. Use either the first conditional matrix or the second conditional matrix as the target conditional matrix. Step 2: Add noise only to the unknown regions in the target condition matrix, while keeping the known obstacle regions unchanged, construct the forward process, and obtain the vector field corresponding to the forward process; Step 3: Using a U-Net structured neural network to start from the noise distribution, introduce reverse block learning and predict the vector field, and gradually reconstruct the complete initial radio map, thus constructing the reverse process; Step 4: Project the first or second conditional guidance tensor in the initial radio map into an embedded representation through the extractor, and inject it into the intermediate layer of U-Net via a multi-head attention mechanism to generate the radio map; thus obtaining the generative flow model RRFlow. Step 5: Train RRFlow by minimizing the flow matching loss function and using the AdamW optimizer to obtain the trained RRFlow.

2. The method for rapid generation of radio maps based on flow matching as described in claim 1, characterized in that, In step one, the first condition matrix in the target condition matrix M Or the second condition matrix for: ； ； In the formula, The first condition guides the tensor. ; This serves as the guiding tensor for the second condition. ; Let k be the channel index in the conditional guidance tensor, N be the number of grids, and (i,j) be the position coordinates on the grid.

3. The method for rapid generation of radio maps based on flow matching as described in claim 2, characterized in that, In step one, the first conditional guiding tensor is: ； The second conditional guiding tensor is: ； In the formula, This is a concatenation function; It is a sparse sampling matrix in grayscale image format. A static obstacle matrix in grayscale image format. A dynamic obstacle matrix in grayscale image format. This is a base station location matrix in grayscale image format.

4. The method for rapid generation of radio maps based on flow matching as described in claim 3, characterized in that, In step one, the method for processing grayscale image format is as follows: S1, the original sparse sampling matrix pass Normalization mapping to The interval is processed into a sparse sampling matrix in grayscale image format. The processing method is as follows: ； In the formula, This is the minimum value of the original sparse sampling matrix; This represents the maximum value of the original sparse sampling matrix; S2, the original static obstacle matrix Static obstacle matrix processed into grayscale image format using binary encoding ;in, Static obstacle elements , indicates that there is a static obstacle at (i,j); , indicates that there is no static obstacle at (i,j); S3, the original dynamic obstacle matrix Dynamic obstacle matrix processed into grayscale image format using binary encoding ;in, Dynamic obstacle elements in , indicating that there is a dynamic obstacle at (i,j); , indicating that there is no dynamic obstacle at (i,j); S4, the original base station location matrix Base station location matrix processed into grayscale image format using binary encoding ;in, Base station location elements , indicating that a base station exists at (i,j); , indicating that there is no base station at (i,j).

5. The method for rapid generation of radio maps based on flow matching as described in claim 4, characterized in that, In step two, the method for constructing the forward process is as follows: ； ； In the formula, This represents the t-th step noise data obtained by combining the t-th step noise data with the target condition matrix M during the forward process, where t is the time step; the t-th step noise data express With Gaussian noise The noise data obtained at step t by element-wise multiplication; To represent the first step in the forward process The first step of the noise data is obtained by combining the noise data with the target condition matrix M. Step-by-step noisy data, It represents the derivative of the total number of steps in time, and in discrete time it represents the reciprocal of the total number of steps from the target distribution to the initial sample distribution; This represents element-wise multiplication. This represents the unknown region in the objective condition matrix. , represented as in the matrix All elements are 1.

6. The method for rapid generation of radio maps based on flow matching as described in claim 5, characterized in that, In step two, the vector field corresponding to the forward process is: ； In the formula, Given the target distribution during the forward process, the data distribution at step t is... Given the target distribution and the initial sample distribution during the forward process, the data distribution at step t is... For the target distribution, For the initial sample distribution, Let be the data distribution at step t.

7. The method for rapid generation of radio maps based on flow matching as described in claim 6, characterized in that, In step three, the reverse block is: ； In the formula, For the first step in the reverse process The t-dt-th step noise data is obtained by combining the step noise data with the target condition matrix M; This represents the noise data at step t-dt without incorporating the target condition matrix M during the reverse process.

8. The method for rapid generation of radio maps based on flow matching as described in claim 7, characterized in that, In step three, the method for constructing the reverse process is as follows: ； In the formula, For the first step in the reverse process The t-th step noise data is obtained by combining the initial noise data with the target condition matrix M. Let be the vector field that the U-Net structure neural network learns and predicts at step t.

9. The method for rapid generation of radio maps based on flow matching as described in claim 8, characterized in that, In step four, the method for injecting the multi-head attention mechanism into the intermediate layers of U-Net is as follows: ； In the formula, To extract attention from the first or second condition-guided tensor in the initial radio map using a neural network with a U-Net structure. , and These are the queries, keys, and values ​​obtained by projecting the intermediate feature maps of a U-Net-structured neural network. The standardized exponential function transforms any set of numbers into a probability distribution through exponential operations and normalization; T is the transpose; d is the query term. s and keys The dimension is used to scale the dot product.

10. The method for rapid generation of radio maps based on flow matching as described in claim 9, characterized in that, In step five, the flow matching loss function is: ； In the formula, To match the loss function value to the flow, For the parameters of the U-Net neural network structure, This indicates the expectation of a predefined function; This indicates that, in a known reverse process, t follows a uniform distribution in [0,1]. ; For the forward process given the intermediate hidden variable z obey distributed, Given the latent variable z The distribution, Let the hidden variable z follow The distribution, Let z be the distribution of the latent variable. Let z be the data distribution at step t given the latent variable z in the forward process, where z is the latent variable. Connect... and .

Images