A method for generating a radio map in a complex environment based on a physically constrained latent diffusion model and an electronic device

Abstract

This invention relates to the interdisciplinary fields of wireless communication network optimization, electromagnetic environment perception, and artificial intelligence computer vision. Specifically, it is a method and electronic device for generating radio maps of complex environments based on a physical constraint latent diffusion model, comprising the following steps: S1, decoupling and constructing a shadow fading gradient map; S2, predicting the shadow fading gradient distribution; S3, compressing the map to a low-dimensional latent space; S4, training the physical constraint latent diffusion model; and S5, inferring and generating the radio map. This invention constructs a hybrid modeling framework of physical prior guidance and latent space generation, introducing a shadow fading gradient map as a strong physical constraint, effectively eliminating the wall-penetrating effect and false signal artifacts. By using a VAE to map high-dimensional data to a low-dimensional latent space for diffusion denoising, it significantly reduces computational resource consumption and GPU memory usage. It achieves the effect of significantly improving the quality of radio map generation and accurately identifying the signal propagation characteristics of complex environments under sparse sampling.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of wireless communication network optimization, electromagnetic environment perception, and artificial intelligence computer vision, specifically a method and electronic device for generating radio maps of complex environments based on a potential diffusion model with physical constraints. Background Technology

[0002] With the full commercialization of fifth-generation mobile communication (5G) and the commencement of research on sixth-generation mobile communication (6G), the deployment density and frequency band height of wireless networks are constantly increasing, and the electromagnetic environment is becoming increasingly complex. As a data structure that describes the spatial distribution of received signal strength, signal-to-interference-plus-noise ratio, or other channel parameters within a specific geographical area, building high-precision radio maps is crucial for network planning, interference management, dynamic scheduling of spectrum resources, and fingerprint-based positioning services. However, constructing high-precision radio maps in complex urban environments faces enormous challenges: on the one hand, data acquisition is costly. Traditional radio map construction relies on extensive road surveys, and conducting dense grid-based measurements in vast urban areas is time-consuming and labor-intensive. Furthermore, it is subject to restrictions such as traffic conditions and restrictions on access to private areas, resulting in extremely sparse data (sampling rates are typically below 1%, or even around 0.5%). On the other hand, the electromagnetic propagation environment is complex. Unlike indoor environments, the environment is vast and contains dense buildings, vegetation, vehicles, and other scattering bodies. Electromagnetic waves undergo complex processes of direct reflection, reflection, diffraction, and scattering during propagation. In particular, the "shadow fading" effect, which causes a sharp drop in signal due to buildings blocking the line-of-sight path, and the "street canyon effect," which causes signals to propagate along street waveguides, result in highly nonlinear and non-stationary spatial distribution of signals. Furthermore, existing construction methods have significant limitations: interpolation methods (such as Kriging and IDW) are based on spatial correlation assumptions and cannot detect building obstructions, resulting in overly smooth predictions that severely violate physical principles; while ray tracing methods offer high accuracy, their computational complexity increases exponentially with the number of scene faces and ray bounces, failing to meet real-time or dynamic environment update requirements; and traditional deep learning methods, such as convolutional neural networks or generative adversarial networks, which have become popular in recent years, typically treat this problem as a simple image inpainting task, often only learning the statistical texture distribution of signal intensity while ignoring the underlying physical mechanisms of electromagnetic wave propagation, leading to poor model generalization and the generation of artifacts such as predicting strong signals in building shadow areas. Therefore, how to effectively integrate the physical knowledge of electromagnetic wave propagation into powerful deep generative models to achieve high-precision radio map reconstruction that conforms to physical laws while ensuring computational efficiency is a key technical problem that urgently needs to be solved by academia and industry. Summary of the Invention

[0003] The purpose of this invention is to overcome the problems of poor physical interpretability, weak generalization ability and low computational efficiency in the existing technology when generating radio maps for complex environments, and to provide a method and electronic device for generating radio maps for complex environments based on a potential diffusion model with physical constraints.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for generating radio maps of complex environments based on a physical constraint-based potential diffusion model includes the following steps:

[0006] S1. Decoupling and Constructing a Shadow Fading Gradient Map: Since the signal strength in the radio map is coupled with deterministic path loss that smoothly attenuates with distance and nonlinear shadow fading caused by obstacle occlusion, this invention, based on the link budget principle, models the signal strength at the receiving end as a superposition mapping of deterministic and random components. Specifically, firstly, by combining the acquired scene data and the free-space propagation model, the free-space received signal strength representing the smooth attenuation characteristics is calculated; then, by decoupling it from the ground truth of the radio map, the shadow fading distribution map representing the nonlinear occlusion effect is extracted; finally, the shadow fading field is processed using a two-dimensional spatial gradient operator to extract a shadow fading gradient map that can accurately characterize building boundary information.

[0007] S2. Predicting the shadow fading gradient distribution: Construct a physical prior prediction network, taking a binarized or height-included building floor plan and a sparse receiver signal strength map as input, and the shadow fading gradient map obtained in step S1 as a supervision label, train the network to predict the shadow fading gradient distribution of the entire region, and finally output the predicted shadow fading gradient map.

[0008] S3. Compress the map to a low-dimensional latent space: Construct and train a variational autoencoder (VAE) model to uniformly map the high-dimensional radio map, building map, sparse receiver signal strength map, and the predicted shadow fading gradient map obtained in step S2 to a low-dimensional latent feature space. After VAE encoding, extract the latent features corresponding to the radio map as the latent target features to be predicted; at the same time, extract the latent features corresponding to the building map, sparse receiver signal strength map, and predicted shadow fading gradient map, and define them uniformly as latent conditional features; after training is completed, freeze the VAE model parameters for use in subsequent stages.

[0009] S4. Training the physical constraint latent diffusion model: Construct a conditional diffusion model that runs in the latent space. The latent target features obtained in step S3 are forward diffused to obtain a random noise that conforms to the isotropic standard normal distribution. Combined with the latent conditional features obtained in step S3, the predicted latent target features are obtained through a reverse generation process. Finally, after passing through the VAE decoder, the predicted radio map is obtained.

[0010] S5. Inference to Generate Radio Map: Obtain building maps and sparse receiver signal strength map data for the area to be tested. Perform inference based on the methods in steps S1-S4 and output the final high-resolution radio map. This includes: obtaining the predicted shadow fading gradient map of the entire area through the physical prior prediction network pre-trained in step S2; encoding the building maps, sparse receiver signal strength map data, and predicted shadow fading gradient map into latent conditional features through the VAE encoder frozen in step S3; sampling random Gaussian noise in the latent space and inputting iterative denoising into the conditional diffusion model constructed in step S4; inputting the denoised latent conditional features into the VAE decoder and outputting the final high-resolution radio map.

[0011] Preferably, the specific process of decoupling and constructing the shadow fading gradient map in step S1 includes:

[0012] S11. Establish a free-space propagation model: Grid the entire region and define the coordinates of any grid point within that region as... Based on the position coordinates of the transmitting end Transmission power Signal frequency and path loss index Calculate any location within the area to be measured. Free space received signal strength After calculating the signal strength of all grid points within the entire region, the signal strength of each grid point will be... By mapping and arranging the signals according to their spatial relationships, a free-space loss map corresponding to the region is constructed. This map intuitively presents the spatial intensity distribution characteristics of radio signals throughout the region under ideal conditions without considering building obstruction and multipath effects, providing a basic analytical reference for subsequent deep learning predictions combined with physical constraints.

[0013] S12. Obtain the shadow fading distribution map: Obtain the radio map truth value. The difference between the received signal strength in free space and the signal strength obtained in step S11 is calculated to obtain the shadow fading distribution map caused by the urban environment. ;

[0014] S13. Obtain the shadow fading gradient map: Apply the two-dimensional spatial gradient operator to the shadow fading distribution map obtained in step S12. The resulting magnitude matrix is ​​the shadow fading gradient map. .

[0015] Preferably, the free space received signal strength in step S11 Represented as:

[0016]

[0017] in, Represents grid points The Euclidean distance to the transmitter; Representation and frequency The path loss constant at the relevant reference distance can be expressed as:

[0018]

[0019] in, At the speed of light, For the transmitting antenna gain at the transmitting end, This represents the antenna gain at the receiving end.

[0020] Preferably, the shadow fading gradient map in step S13 Represented as:

[0021]

[0022] in, This represents the gradient operator in two-dimensional space, specifically expressed as follows:

[0023] .

[0024] The gradient plot Mathematically, it characterizes the "discontinuity" or "abrupt rate" of the spatial distribution of electromagnetic field energy, accurately delineating the boundaries of all buildings that produce diffraction and blocking effects on signal propagation, thus constituting the strong physical prior of this method.

[0025] Preferably, the physical prior prediction network in step S2, which predicts the shadow fading gradient distribution, is a deep convolutional neural network based on an encoder-decoder structure. The deep convolutional neural network adopts an improved Res-UNet architecture. The input layer receives the channel splicing tensor of the building map and the sparse receiver signal intensity map. The encoder part uses a residual connection module to extract multi-scale environmental features. The decoder part fuses shallow geometric features through skip connections. The output layer outputs the predicted shadow fading gradient map through 1×1 convolution. The loss function is a weighted sum of L1 norm loss and gradient difference loss.

[0026] Preferably, step S3, compressing the map into a VAE in a low-dimensional latent space, includes an encoder. and decoder In the encoder and decoder Gaussian sampling layers are set between the layers; during the training of the VAE model, the evidence lower bound ELBO objective function is optimized, which includes a reconstruction error term and a KL divergence regularization term; through training, the VAE model extracts various complex propagation patterns in the radio map, including direct paths, shadow areas, and multipath fading areas, and encodes them into compact latent features of the real radio map. .

[0027] Preferably, step S4, which generates the conditional diffusion model in the potential space physics constraints, employs a U-Net denoising network based on a cross-attention mechanism. Specifically, it includes:

[0028] Forward diffusion process: The latent features of the encoded real radio map obtained in step S3 are then... Gaussian noise is gradually added until the time step is reached. Random noise transformed into an isotropic standard normal distribution , where any time step Noise potential features It can be represented as:

[0029]

[0030] in, , Indicates the cumulative multiplication coefficient. The sampled standard Gaussian noise, It is the identity matrix; The preset variance scheduling parameter has a value range of [value range missing]. As time goes by The increase in usually shows a linear or cosine trend; as , , making It approximately follows a standard normal distribution ;

[0031] Reverse generation process: from random noise Begin, at every arbitrary time step Denoising network Receive current noise potential features Time step embedding vector and condition vector Calculate the latent features of the previous time step :

[0032]

[0033] in, It is a joint latent conditional feature that includes the shadow fading gradient map, the building map, and the sparse receiver signal strength map; It is the noise component predicted by the conditional diffusion model in step S4; To control the generated random variance term, it is usually taken as... ; For random noise term, when hour ;

[0034] Conditional injection mechanism: condition vector It is composed of latent features from the predicted shadow fading gradient map, latent features from the sparse receiver signal strength map, and latent features from the building map, and is then processed by a denoising network. The intermediate layer is injected into the denoising network by adding channel splicing. This guides the denoising direction towards convergence with the signal distribution that conforms to physical laws.

[0035] Preferably, the specific process of generating the radio map through reasoning in step S5 includes:

[0036] S51. Data Acquisition: Acquire a building map and a sparse receiver signal strength map for the entire area;

[0037] S52. Obtain the predicted shadow fading gradient map: Based on the building map and sparse receiver signal strength map obtained in step S51, obtain the predicted shadow fading gradient map of the entire region through the physical prior prediction network pre-trained in step S2.

[0038] S53. Obtain latent conditional features: Encode the building map, sparse receiver signal strength map and predicted shadow fading gradient map into latent conditional features using the VAE encoder frozen in step S3.

[0039] S54. Iterative denoising: Random Gaussian noise is sampled in the latent space and input together with the latent conditional features obtained in step S53 into the conditional diffusion model constructed in step S4 for iterative denoising to obtain the denoised latent target features.

[0040] S55, Output Radio Map: Input the denoised potential target features into the VAE decoder to obtain the final high-resolution radio map.

[0041] An electronic device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement any of the steps in the above-described method for generating radio maps of complex environments based on a potential diffusion model with physical constraints.

[0042] The beneficial effects of this invention are:

[0043] This invention innovatively constructs a hybrid modeling framework of "physical prior guidance + latent space generation", and introduces the "shadow fading gradient map" that accurately characterizes the diffraction and shading effects of electromagnetic waves at the edge of buildings as a strong physical constraint. This mechanism forces the generation process to strictly follow the physical boundary of electromagnetic propagation, thereby significantly breaking through the accuracy bottleneck of traditional methods under extremely sparse sampling conditions and effectively eliminating the "wall-penetrating effect" and false signal artifacts.

[0044] This invention further utilizes a variational autoencoder (VAE) to map high-dimensional panoramic data to a low-dimensional latent space for diffusion denoising. Compared with traditional pixel-level generation models, it significantly reduces computational resource consumption and GPU memory usage while ensuring high resolution and high fidelity reconstruction quality, ultimately achieving a synergistic improvement in physical interpretability, generation quality, and computational efficiency. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the overall architecture of Embodiment 1 of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the calculation principle of the shadow fading gradient map in Embodiment 1 of the present invention;

[0047] Figure 3 This is a schematic diagram of the network structure of the physical prior prediction network in Embodiment 1 of the present invention;

[0048] Figure 4 This is a schematic diagram of the training process of the conditional latent diffusion model in Embodiment 1 of the present invention;

[0049] Figure 5 This is a schematic diagram illustrating the visualization comparison of the method of the present invention with the PMNet method and the radio map truth prediction results in Embodiment 2 of the present invention. Detailed Implementation

[0050] Example 1

[0051] The following is a further explanation of the present invention in conjunction with specific embodiments, such as... Figure 1 As shown, this embodiment is a method for generating radio maps in complex environments based on a potential diffusion model with physical constraints, which includes the following steps:

[0052] First, we construct the system model and define the scenario:

[0053] This embodiment uses the publicly available dataset RadioMapSeer as its data foundation to verify the model's generalization performance in standard urban scenarios. This dataset contains 701 building layout maps of real cities (such as Ankara, Berlin, London, etc.) extracted from OpenStreetMap. First, the scene is digitized, and the target area is set as... A two-dimensional grid of pixels, with a resolution of 1 meter per pixel. Regarding the input data definition, the building map (...) The system uses a 2.5D map with height information provided by RadioMapSeer, where pixel values ​​represent building heights; the transmitter information is based on coordinates provided in the dataset. Confirmed, signal parameters set to 5.9GHz band, transmit power 23dBm; sparse receiver signal strength graph. It is generated by applying a mask to the original complete radio map. To simulate a real, extremely sparse sampling environment, this embodiment strictly sets the sampling rate to approximately 0.5%, meaning that only about 328 valid measurement points are retained out of the 65,536 grids in the entire map. The final output target is the true value of the radio map (…). This refers to a high-precision received signal strength distribution map generated by the IRT (Intelligent Ray Tracing) algorithm in the WinProp software.

[0054] S1. Decoupling and Constructing the Shadow Fading Gradient Map: A physical prior extraction mechanism is constructed. Using the transmitter location information and free-space propagation model provided in the dataset, the received signal strength in free space is calculated. This is then decoupled from the ground truth radio map to extract the shadow fading distribution map and calculate the shadow fading gradient map. The specific process includes:

[0055] S11. Establish a free-space propagation model: such as... Figure 2 As shown, the entire region is gridded, and the coordinates of any grid point within the region are defined as follows: Based on the position coordinates of the transmitting end Transmission power Signal frequency and path loss index (In this embodiment, it can be taken as follows) The free space loss algorithm is used to calculate the value at any location within the entire region. Free space received signal strength ;

[0056] Free space received signal strength Represented as:

[0057]

[0058] in, Represents grid points The Euclidean distance to the transmitter; Representation and frequency The path loss constant at the relevant reference distance can be expressed as:

[0059]

[0060] in, At the speed of light, For the transmitting antenna gain at the transmitting end, This represents the antenna gain at the receiving end.

[0061] After completing the signal strength calculation for all grid points in the entire region, the signal strength of each grid point will be... By mapping and arranging the signals according to their spatial relationships, a free-space loss map corresponding to the region is constructed. This map visually presents the spatial intensity distribution characteristics of radio signals throughout the region under ideal conditions, without considering building obstruction and multipath effects, providing a basic analytical reference for subsequent deep learning predictions combined with physical constraints.

[0062] S12. Obtain the shadow fading distribution map: Obtain the radio map truth value. The difference between the received signal strength in free space and the signal strength obtained in step S11 is calculated to obtain the shadow fading distribution map caused by the urban environment. ;

[0063] S13. Obtain the shadow fading gradient map: Apply the two-dimensional spatial gradient operator to the shadow fading distribution map obtained in step S12. The resulting magnitude matrix is ​​the shadow fading gradient map. ;

[0064] Shadow fading gradient map Represented as:

[0065]

[0066] in, This represents the gradient operator in two-dimensional space, specifically expressed as follows:

[0067] .

[0068] The gradient plot Mathematically, it characterizes the "discontinuity" or "abrupt rate" of the spatial distribution of electromagnetic field energy, accurately delineating the boundaries of all buildings that produce diffraction and blocking effects on signal propagation, thus constituting the strong physical prior of this method.

[0069] S2. Predicting Shadow Fading Gradient Distribution: Since the gradient map cannot be directly calculated during the inference phase due to the lack of ground truth values, this step designs and trains a deep convolutional neural network to construct a physical prior prediction network. The aim is to establish a geometric domain of the environment and sparse observation domain To the physical gradient domain Nonlinear mapping relationship:

[0070]

[0071] in, This represents the predicted shadow fading gradient map output by the network. These are the network parameters. The training objective of this network is to minimize the predicted gradient field. With the real gradient field By studying the distribution differences between them, supervised learning enables the network to grasp the intrinsic physical relationship between the "building geometry" and the "electromagnetic field mutation boundary", thus providing high-precision physical guidance for the subsequent generation process.

[0072] The specific structure of the network is as follows: Figure 3 As shown; this physical prior prediction network is based on an encoder-decoder structure and adopts an improved Res-UNet architecture, combining the advantages of residual learning and U-Net; the network input layer receives a tensor with 2 channels, containing a normalized building map. Sparse receiver signal strength diagram The encoder consists of 4 ResBlocks, each ResBlock containing two... The code employs convolutional layers, batch normalization (BN) layers, and ReLU activation functions, introducing residual connections and downsampling via max pooling layers. The bottleneck layer uses a 1024-channel convolutional layer to extract highly abstract environmental semantic features. The decoder consists of four upsampling modules, each expanding the feature map size through bilinear interpolation and establishing skip connections with the corresponding feature maps from the encoder layer, followed by feature fusion through convolutional layers. The final output layer... Convolution output single-channel predicted shadow fading gradient map In terms of training strategy, L1Loss is used for supervised training.

[0073] S3. Compress the map to a low-dimensional latent space: In order to handle large-scale maps and reduce the GPU memory overhead of the diffusion model, this step uses VAE to compress the data to the latent space; construct and train a variational autoencoder (VAE) model to map the high-dimensional radio map, building map, sparse receiver signal strength map and shadow fading gradient map to low-dimensional latent feature vectors, and obtain the dimensionality-reduced shadow fading gradient map; freeze the VAE model parameters after training.

[0074] VAE includes encoder and decoder In the encoder and decoder A Gaussian sampling layer is set between them; encoder It contains 3 downsampling convolutional blocks, each consisting of a convolutional layer with a stride of 2 and a residual block, reducing the input resolution to 1. Image compression to The feature map was expanded to have 4 channels. Latent features were obtained through the encoder. Output mean and logarithmic variance And utilize reparameterization techniques (in The decoder obtains the output through sampling. Structure and Encoder Symmetrical, containing 3 upsampling blocks, responsible for upsampling latent features The original pixel space is reconstructed. The VAE model is pre-trained using the complete RadioMapSeer training set. During the training process, the ELBO objective function, representing the lower bound of evidence, is optimized to ensure the continuity of the latent space and the quality of reconstruction.

[0075]

[0076] in, is the total loss function of the variational autoencoder (i.e., the negative lower bound of evidence). The input is high-dimensional radio map data; This is the compressed low-dimensional latent feature vector (latent feature); The posterior probability distribution fitted to the encoder; This is the likelihood probability distribution (i.e., the reconstruction generation distribution) corresponding to the decoder. The predefined prior distribution of latent features (usually set as a standard normal distribution) ); This represents the expectation under the posterior distribution predicted by the encoder, with the first term being the reconstruction error; This represents the Kullback-Leibler divergence, with the second term being a regularization term to ensure that the latent variables... The prior follows a standard normal distribution, thus providing a regular probability space for the generation process of the diffusion model.

[0077] The VAE model is trained to extract various complex propagation patterns from radio maps, including direct paths, shadow areas, and multipath fading areas, and encodes them into a compact distribution of latent features.

[0078] S4. Training the physical constraint-based latent diffusion model: such as... Figure 4 As shown, a conditional diffusion model operating in the latent space is constructed. The dimensionality-reduced shadow fading gradient map, the dimensionality-reduced sparse receiver signal strength map, and the building map obtained in step S3 are used as latent conditional features. The latent features of the radio map are generated through the back diffusion process.

[0079] Step S4 generates a conditional diffusion model for the potential space physics constraints using a U-Net denoising network based on a cross-attention mechanism. Specifically, it includes:

[0080] First, data preparation for the potential space is performed using a frozen encoder. Mapping radio maps to potential target features Mapping the predicted shadow fading gradient map to physical latent condition features Mapping building maps to geometric latent condition features Mapping sparse receiver signal intensity maps to latent observation condition features These three latent conditional features are then concatenated along the channel dimension to form a joint latent conditional feature. .

[0081] Forward diffusion process: The encoded real radio map features obtained in step S3 are used to generate potential target features. Gaussian noise is gradually added until the time step is reached. Random noise transformed into an isotropic standard normal distribution , where any time step Noise potential features It can be represented as:

[0082]

[0083] in, , Indicates the cumulative multiplication coefficient. The sampled standard Gaussian noise, It is the identity matrix; The preset variance scheduling parameter has a value range of [value range missing]. As time goes by The increase in usually shows a linear or cosine trend; as , , making It approximately follows a standard normal distribution .

[0084] In the forward diffusion process of the conditional diffusion model, according to the preset linear variance scheduling table ( The total time steps increase linearly from 0.0001 to 0.02. ), at any time Noisy latent features It can be obtained directly through sampling using a formula based on the cumulative variance coefficient.

[0085] Reverse generation process: U-Net denoising network Receive the noise characteristics at the current moment And time steps are implemented through MLP Encoding injection network; joint condition By injecting through channel splicing, the network can dynamically adjust the receptive field weights based on the highlighted areas of the gradient map.

[0086] From random noise Begin, at every arbitrary time step Denoising network Receive current noise potential features Time step embedding vector and condition vector Calculate the latent features of the previous time step :

[0087]

[0088] in, It is a joint latent conditional feature that includes the shadow fading gradient map, the building map, and the sparse receiver signal strength map; It is the noise component predicted by the conditional diffusion model in step S4; To control the generated random variance term, it is usually taken as... ; For random noise term, when hour .

[0089] Conditional injection mechanism: condition vector It is composed of latent features from the predicted shadow fading gradient map, latent features from the sparse receiver signal strength map, and latent features from the building map, and is then processed by a denoising network. The intermediate layer is injected into the denoising network by adding channel splicing. This guides the denoising direction towards convergence with the signal distribution that conforms to physical laws.

[0090] Training a conditional diffusion model is essentially learning a conditional noise predictor. Its optimization objective is to minimize the time step Latent variables and physical conditions Weighted mean square error:

[0091]

[0092] in, This is the training loss function for the conditional latent diffusion model; This is real Gaussian noise sampled from a standard normal distribution; This represents the expectation of all random variables. Condition vector. The predicted shadow fading gradient features are explicitly included. During the reverse generation process, these gradient features act as a "physical guide," directing the denoising network. When recovering signal texture from Gaussian noise, the wavefront evolution of the constrained signal is strictly aligned with the physical boundaries of the gradient, thereby eliminating non-physical artifacts and blurring.

[0093] S5. Inference to Generate Radio Map: Obtain a building map and sparse receiver signal strength map for the entire area, perform inference based on the methods in steps S1-S4, and output the final high-resolution radio map. This includes the following steps:

[0094] S51. Data Acquisition: Acquire a building map and a sparse receiver signal strength map for the entire area;

[0095] S52. Obtain the predicted shadow fading gradient map: Based on the building map and sparse receiver signal strength map obtained in step S51, obtain the predicted shadow fading gradient map of the entire region through the physical prior prediction network pre-trained in step S2.

[0096] S53. Obtain latent conditional features: Encode the building map, sparse receiver signal strength map and predicted shadow fading gradient map into latent conditional features using the VAE encoder frozen in step S3.

[0097] S54. Iterative denoising: Random Gaussian noise is sampled in the latent space and input together with the latent conditional features obtained in step S53 into the conditional diffusion model constructed in step S4 for iterative denoising to obtain the denoised latent target features.

[0098] S55, Output Radio Map: Input the denoised potential target features into the VAE decoder to obtain the final high-resolution radio map.

[0099] Example 2

[0100] To fully verify the effectiveness of this invention in complex environments, this embodiment conducted a comparative experiment based on the publicly available dataset RadioMapSeer. 701 city maps and their corresponding WinProp ray tracing simulation data (5.9GHz band) were selected, and the training, validation, and test sets were divided in a 7:2:1 ratio. The sampling rate was strictly set to 0.5% (i.e., only about 328 observations were retained out of the 65,536 pixels in the entire image) to simulate an extremely sparse road testing scenario. Existing models such as PPNet, Agile, and PMNet, as well as a common diffusion model without physical constraints, were selected for comparison. RMSE (Root Mean Square Error), NMSE (Normalized Root Mean Square Error), SSIM (Structural Similarity Index), and PSNR (Peak Signal-to-Noise Ratio) were used as evaluation metrics, and the results are shown in Table 1.

[0101] Table 1. Experimental results of our method and other deep learning methods in complex urban scenarios.

[0102] method RMSE NMSE SSIM PSNR PPNet 0.0507 - - - Agile 0.0451 - - - PMNet 0.038 - - - diffusion model 0.016 0.0020 0.97 45.04 This method 0.015 0.0016 0.98 45.84

[0103] As shown in Table 1, the quantitative results first reveal the significant advantages of the generative diffusion model compared to traditional methods: PPNet, Agile, and PMNet have high reconstruction errors under sparse sampling (RMSE of 0.0507, 0.0451, and 0.038, respectively), making it difficult to effectively capture the complex propagation laws of electromagnetic waves; while the basic diffusion model, with its powerful generative ability, directly reduces the RMSE to 0.016 and improves the SSIM to 0.97, indicating that the diffusion model has significant performance advantages in handling sparse data reconstruction tasks.

[0104] Building upon this foundation, this method further overcomes performance bottlenecks by introducing a shadow fading gradient map as a strong physical constraint, achieving even better results: compared to the basic diffusion model, the RMSE of this method converges further to 0.015, the NMSE decreases from 0.0020 to 0.0016 (a 20% reduction in relative error), the SSIM increases from 0.97 to 0.98, and the PSNR improves from 45.04 to 45.84. These data strongly demonstrate that, in addition to the superior generation capabilities of the diffusion model itself, the injection of physical priors effectively eliminates "ghost signals," such as... Figure 5 As shown, where Figure 5 (a) shows the prediction results of the PMNet method. Figure 5 (b) shows the prediction results of this method. Figure 5 (c) represents the true value of the radio map; Figure 5(a) The area marked by the red box in the PMNet prediction results is used as a reference. It can be seen that the PMNet method exhibits significant distortion when restoring the small buildings and their corresponding shadow details in this area. In contrast, this method ( Figure 5 (b) Within the same corresponding area, it can reproduce the specific form of small buildings with extremely high accuracy, and the sense of layering and direction of the shadow distribution is consistent with the true value. Figure 5 (c) A perfect match. According to Figure 5 The comparison shows that the present invention is highly consistent with the true value in the restoration of texture details, building edges and shadow contours, thus achieving a dual improvement in accuracy and physical interpretability.

[0105] The above description is merely a further explanation of the present invention in conjunction with specific embodiments. All descriptions made do not imply any limitation on the scope of protection of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology 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 generating radio maps of complex environments based on a potential diffusion model with physical constraints, characterized in that, Includes the following steps: S1. Decoupling and constructing the shadow fading gradient map: Construct a physical prior extraction mechanism, take the transmitter location information and the size of the entire area as input, and use the free space propagation model to calculate the free space loss map; then input the radio map truth value to decouple and obtain the shadow fading distribution map, and finally output the shadow fading gradient map through the spatial gradient operator. S2. Predict the shadow fading gradient distribution: Construct a physical prior prediction network, taking a binarized or height-included building floor plan and a sparse receiver signal strength map as input, and the shadow fading gradient map obtained in step S1 as the supervision label, train the network to predict the shadow fading gradient distribution of the entire region, and output the predicted shadow fading gradient map. S3. Compress the map to a low-dimensional latent space: Construct and train a variational autoencoder (VAE) model to uniformly map the high-dimensional radio map, building map, sparse receiver signal strength map, and shadow fading gradient map predicted in step S2 to a low-dimensional latent feature space; after VAE encoding, extract the latent features corresponding to the radio map as the latent target features to be predicted. Simultaneously, the latent features corresponding to the building map, sparse receiver signal strength map, and predicted shadow fading gradient map are extracted and uniformly defined as latent conditional features; Once training is complete, freeze the VAE model parameters; S4. Training the physical constraint latent diffusion model: Construct a conditional diffusion model that runs in the latent space. The latent target features obtained in step S3 are forward diffused to obtain a random noise that conforms to the isotropic standard normal distribution. Combined with the latent conditional features obtained in step S3, the predicted latent target features are obtained through a reverse generation process. Finally, after passing through the VAE decoder, the predicted radio map is obtained. S5. Inference to generate a radio map: Obtain a building map and a sparse receiver signal strength map of the entire area, perform inference based on the methods in steps S1-S4, and output the final high-resolution radio map.

2. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 1, characterized in that: The specific process of decoupling and constructing the shadow fading gradient map in step S1 includes: S11. Establish a free-space propagation model: based on the position coordinates of the transmitter. Transmission power Signal frequency and path loss index Calculate any location within the entire region Free space received signal strength ; S12. Obtain the shadow fading distribution map: Obtain the radio map truth value. The difference between the received signal strength in free space and the signal strength obtained in step S11 is calculated to obtain the shadow fading distribution map caused by the urban environment. ; S13. Obtain the shadow fading gradient map: Apply the two-dimensional spatial gradient operator to the shadow fading distribution map obtained in step S12. The resulting magnitude matrix is ​​the shadow fading gradient map. .

3. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 2, characterized in that: The free space received signal strength in step S11 Represented as: in, Represents grid points The Euclidean distance to the transmitter; Representation and frequency The path loss constant at the relevant reference distance can be expressed as: in, At the speed of light, For the transmitting antenna gain at the transmitting end, This represents the antenna gain at the receiving end.

4. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 3, characterized in that: The shadow fading gradient map in step S13 Represented as: in, This represents the gradient operator in two-dimensional space, specifically expressed as follows: 。 5. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 4, characterized in that: The physical prior prediction network in step S2, which predicts the shadow fading gradient distribution, is a deep convolutional neural network based on an encoder-decoder structure. The deep convolutional neural network adopts an improved Res-UNet architecture. The input layer receives the channel splicing tensor of the building map and the sparse receiver signal intensity map. The encoder part uses a residual connection module to extract multi-scale environmental features. The decoder part fuses shallow geometric features through skip connections. The output layer outputs the predicted shadow fading gradient map through 1×1 convolution. The loss function is a weighted sum of L1 norm loss and gradient difference loss.

6. The complex environment radio map generation system based on a physical constraint potential diffusion model according to claim 5, characterized in that: Step S3, compressing the map into a low-dimensional latent space VAE, includes an encoder. and decoder In the encoder and decoder A Gaussian sampling layer is set between the layers; during the training of the VAE model, the evidence lower bound ELBO objective function is optimized, which includes a reconstruction error term and a KL divergence regularization term; through training, the VAE model extracts various complex propagation patterns in the radio map, including direct paths, shadow areas, and multipath fading areas, and encodes them into compact latent features of the real radio map. .

7. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 6, characterized in that: The condition diffusion model generated in step S4 for the potential space physics constraints employs a U-Net denoising network based on a cross-attention mechanism. Specifically, it includes: Forward diffusion process: The latent features of the encoded real radio map obtained in step S3 are then... Gaussian noise is gradually added until the time step is reached. Random noise transformed into an isotropic standard normal distribution , where any time step Noise potential features It can be represented as: in, , Indicates the cumulative multiplication coefficient. The sampled standard Gaussian noise, It is the identity matrix; The preset variance scheduling parameter has a value range of [value range missing]. As time goes by The increase in usually shows a linear or cosine trend; as , , making It approximately follows a standard normal distribution ; Reverse generation process: from random noise Begin, at every arbitrary time step Denoising network Receive current noise potential features Time step embedding vector and condition vector Calculate the latent features of the previous time step : in, It is a joint latent conditional feature that includes shadow fading gradient map, building map and sparse receiver signal strength map; It is the noise component predicted by the conditional diffusion model in step S4; To control the generated random variance term, it is usually taken as... ; For random noise term, when hour ; Conditional injection mechanism: the condition vector It is composed of latent features from the predicted shadow fading gradient map, latent features from the sparse receiver signal strength map, and latent features from the building map, and is then processed by a denoising network. The intermediate layer is injected into the denoising network by adding channel splicing. This guides the denoising direction towards convergence with the signal distribution that conforms to physical laws.

8. The method for generating radio maps of complex environments based on a potential diffusion model with physical constraints according to claim 7, characterized in that: The specific process of generating the radio map in step S5 includes: S51. Data Acquisition: Acquire a building map and a signal strength map of sparse receivers for the entire area; S52. Obtain the predicted shadow fading gradient map: Based on the building map and sparse receiver signal strength map obtained in step S51, obtain the predicted shadow fading gradient map of the entire region through the physical prior prediction network pre-trained in step S2. S53. Obtain latent conditional features: Encode the building map, sparse receiver signal strength map and predicted shadow fading gradient map into latent conditional features using the VAE encoder frozen in step S3. S54. Iterative denoising: Random Gaussian noise is sampled in the latent space, and the random Gaussian noise and the latent conditional features obtained in step S53 are input together into the conditional diffusion model constructed in step S4 for iterative denoising to obtain the denoised latent target features. S55, Output Radio Map: Input the denoised potential target features into the VAE decoder to obtain the final high-resolution radio map.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements any step in the method for generating radio maps of complex environments based on a physical constraint potential diffusion model as described in any one of claims 1-8.

Images