Two-dimensional radio map generation system integrating communication base station and environment information

Abstract

The invention provides a two-dimensional radio map generation system fusing communication base station and environment information, and the system comprises a base station encoder which is used for converting physical parameters of a communication base station in a target region into feature vectors reflecting electromagnetic rebroadcasting features; the environment encoder is used for converting the map of the target area containing the environment information into environment image features reflecting the global features of the target area; the environment image features and the feature vectors are fused to generate joint feature representation; and the radio map decoder is used for generating a radio signal power plane distribution diagram of the target area through a decoding network based on the joint feature representation. The accuracy of the two-dimensional radio map is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of radio signal detection technology, specifically relating to a two-dimensional radio map generation system that integrates communication base station and environmental information. Background Technology

[0002] In the field of integrating communication base station and environmental information to generate two-dimensional radio maps, existing technologies struggle to effectively integrate base station deployment data and environmental elements into a unified model. Traditional methods often process base station parameters (such as transmit power, antenna mode, and site location) and environmental features (such as building distribution, terrain undulation, and land cover) separately, lacking a multi-modal feature fusion mechanism, resulting in an incomplete depiction of the wireless signal propagation environment. Consequently, the generated two-dimensional radio maps have limited accuracy and fail to fully reflect the impact of environmental obstruction and reflection on the actual signal distribution.

[0003] Furthermore, purely data-driven mapping methods may ignore existing propagation patterns, resulting in predictions that sometimes contradict physical expectations; while purely model-driven methods (such as the classical path loss formula) are difficult to adapt to complex and ever-changing real-world environments.

[0004] It is evident that the current approach to integrating multi-source information and knowledge from the field of communication faces technical challenges. The lack of a unified integration framework guided by the laws of communication limits the accuracy and generalization ability of existing radio map generation methods. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the aforementioned background technology and provide a two-dimensional radio map generation system that integrates communication base station and environmental information, thereby effectively improving the accuracy of two-dimensional radio maps.

[0006] The technical solution adopted in this invention is: a two-dimensional radio map generation system that integrates communication base station and environmental information, comprising: A base station encoder is used to convert the physical parameters of communication base stations within a target area into feature vectors that reflect the characteristics of electromagnetic relay. An environment encoder is used to convert a map of a target area containing environmental information into environmental image features that reflect the global characteristics of the target area; and to fuse the environmental image features and feature vectors to generate a joint feature representation. A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation.

[0007] The beneficial effects of this invention are as follows: By processing base station parameters and environmental information in the same system, this invention overcomes the limitations of traditional methods that process the two types of data separately, and fully reflects the comprehensive impact of communication base station deployment and surrounding environment on signal propagation; the system utilizes an encoding-fusion-decoding design, enabling the model to learn environmental features such as occlusion and reflection in a multimodal manner and perform comprehensive analysis in conjunction with base station parameters, thus solving the problem of inaccurate mapping caused by the lack of a multi-source fusion mechanism in existing technologies; and by achieving end-to-end two-dimensional radio map generation under a unified framework, it not only improves prediction accuracy but also ensures the physical rationality and generalization ability of the generated results by integrating knowledge in the propagation domain.

[0008] Furthermore, by introducing rich layers such as land cover type, building height, and altitude, the system can more precisely depict the environmental obstruction and multipath propagation of radio waves, solving the problem of incomplete environmental elements and insufficient prediction accuracy caused by relying solely on simple maps (such as rough plan maps). This combination of environmental information enables multi-angle capture of complex terrain and building distribution in urban areas, allowing the generated two-dimensional radio map to better reflect the actual signal distribution characteristics in scenarios with high-rise buildings and undulating terrain. The rich environmental layers provide a solid data foundation for subsequent fusion, enabling the model to not only grasp the influence of large-scale terrain but also take into account the high-frequency shadowing effect caused by buildings.

[0009] Furthermore, the introduction of the attention module in this invention can adaptively allocate feature weights during the fusion stage, enabling the system to more accurately capture the key contributions of environmental occlusion or changes in base station parameters to signal attenuation, overcoming the drawbacks of simple splicing in existing technologies that leads to mutual interference of multimodal information. Through further convolution or multi-scale processing, the fused features are refined in high-dimensional space, which can more sensitively capture small-scale reflections, shadows, and other factors in complex urban areas, effectively improving the prediction accuracy of radio maps. The phased and multi-step feature extraction and fusion process makes the generated model more robust, and can still efficiently and uniformly represent even when there are large differences between environmental information and base station information.

[0010] Furthermore, the adoption of key base station physical parameters in this invention can more comprehensively reflect the impact of base stations on signal radiation, thereby achieving more accurate modeling in both large-scale path loss and local beam coverage, overcoming the shortcomings of traditional methods that only consider some base station attributes. By simultaneously incorporating carrier frequency, transmit power, downtilt angle, and base station height into the feature vector, the model can better distinguish the differences between high-frequency and low-frequency propagation characteristics and antenna directivity, avoiding blind spots or overly simplified signal estimation. This parameter set enables the system to adapt to different types and configurations of base station deployment scenarios, enhancing the applicability and versatility of the model in actual network planning.

[0011] Furthermore, by injecting random noise into the base station encoder, this invention enables the system to simulate a wider range of real-world scenarios (such as measurement errors or uncertainties) during the training phase, thereby improving the model's tolerance to noise and input disturbances. Compared to purely data-driven methods, this method avoids the problem of the network overfitting to a specific data distribution, ensuring that the generated radio maps maintain high accuracy in different scenarios and significantly enhancing their generalization ability. This random noise mechanism further promotes data diversity, helping the model capture the changing patterns under different propagation environments with the same amount of training data.

[0012] Furthermore, the combination of supervised learning and physical propagation priors in this invention ensures that the model can learn real data features without deviating from known electromagnetic laws, overcoming the risk of physical inconsistencies or inexplicable problems that easily arise from purely data-driven models. The dual approach of empirical and deterministic priors solves the shortcomings of traditional models in terms of large-scale attenuation and small-scale shading effects, enabling the generated radio maps to demonstrate higher accuracy at both macro and micro levels. Through end-to-end training, the various modules of the model achieve collaborative optimization, fully integrating multi-source information and physical knowledge in the complete training pipeline, and realizing in-depth modeling of complex urban environments "from input to output".

[0013] Furthermore, by comparing the theoretical RSRP statistical distribution, this invention effectively supervises the system at the large-scale path loss level, enabling the base station encoder to maintain consistency with classical path loss laws when processing information such as transmit power and carrier frequency. This empirical loss can effectively correct the model's deviation in large-scale attenuation, overcome the defect of pure deep networks being sensitive to some local anomalies, and thus improve the model's adaptability to changes in the macroscopic environment. Strengthening the constraints on the base station encoder output helps the network converge quickly and achieve interpretable physical consistency even with fewer measured samples.

[0014] Furthermore, the auxiliary map generated by the ray tracing model in this invention can reflect details such as buildings, shadows, reflections, and multipath at the microscopic level, enabling the generative network to learn richer high-frequency information. The adversarial training method incorporates physical priors into the iterative process of the generative network, ensuring that the output map is consistent with the real propagation scene at a high level when fusing environmental information and base station parameters, overcoming the limitations of traditional methods in capturing complex occlusions and reflections. By dynamically comparing the model output with "real samples" through the discriminator network, the model's ability to learn shadow boundaries and local features is enhanced, significantly improving the prediction accuracy of two-dimensional radio maps in complex urban areas. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system framework and training process of the present invention; Figure 2This is a schematic diagram of the feature fusion process of the present invention. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions do not constitute a limitation on the present invention.

[0017] Example 1 like Figure 1 As shown, the present invention provides a two-dimensional radio map generation system that integrates communication base station and environmental information, comprising: A base station encoder is used to convert the physical parameters of communication base stations within a target area into feature vectors that reflect the characteristics of electromagnetic relay. An environment encoder is used to convert a map of a target area containing environmental information into environmental image features that reflect the global characteristics of the target area; and to fuse the environmental image features and feature vectors to generate a joint feature representation. A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation.

[0018] The entire system adopts an end-to-end architecture. The inputs are environmental distribution map data and communication base station parameters, and the output is the planar distribution of the predicted signal received power at each location in the target area. The system structure clearly divides each functional module: the environmental encoder and the base station encoder extract two types of input features respectively. The environmental encoder integrates an attention fusion module to achieve the fusion of joint feature representations. Finally, the radio map decoder generates the required radio power map.

[0019] Specifically, the map of the target area containing environmental information includes a map of the distribution of land features, a map of building heights, and a map of ground elevation.

[0020] Environmental information is input into the environmental encoder in the form of a multi-channel rasterized map. The environmental information explicitly includes the following three distribution maps: feature type distribution map, building height distribution map, and ground elevation distribution map. Among them, the feature type distribution map reflects the spatial distribution categories of different features (such as buildings, roads, water bodies, vegetation, etc.) in the target area; the building height distribution map gives the height values ​​of buildings at each location; and the ground elevation distribution map provides the distribution of terrain undulations, i.e., ground elevation.

[0021] The environment encoder preferably employs a convolutional neural network (CNN) structure to extract features from the aforementioned multi-channel environmental distribution map. Through layer-by-layer convolution and pooling operations, the environment encoder extracts multi-scale environmental image feature representations, capturing geographical and architectural environmental patterns that affect wireless propagation within the region. During the extraction process, the environment encoder retains key spatial details and layouts from the input environmental information, enabling the subsequent fusion module to utilize these environmental features to explain signal attenuation and occlusion effects.

[0022] Preferably, this embodiment is applied to the generation of radio maps in complex urban areas. Considering the significant impact of complex urban environments on radio wave propagation, a high-precision environmental map is required. For example... Figure 2 As shown, the environmental information of a residential community can be depicted using various types of maps. The ground elevation distribution map shows the community's altitude distribution, the building height distribution map indicates the building height distribution, and the feature type distribution map marks the terrain environment within the community, with colors distinguished by terrain numbers. The community environment is divided into several major terrain regions, each marked with a corresponding color. These different types of image-based environmental information, together with the aforementioned physical feature vectors of the communication base station, constitute the multimodal input of the environmental encoder.

[0023] Preferably, this embodiment employs a satellite image feature extraction module (autoencoder) to extract ground feature information (such as buildings, vegetation, roads, etc.) from satellite images. The autoencoder, through encoding and decoding processes, converts the satellite remote sensing map into feature representations of different ground feature types, and uses these features to assist in the subsequent generative modeling. Using existing pre-trained models for satellite image ground feature classification tasks, auxiliary labels for ground feature categories are introduced, and a ground feature loss function is established to describe the ground feature information extraction performance of the satellite image feature extraction module. By reducing the ground feature loss, the gradient of the generative model parameters is backpropagated to guide the satellite image extraction module in extracting ground feature information that affects radio propagation.

[0024] Specifically, the physical parameters of the communication base station include location, carrier frequency, antenna transmit power, vertical downtilt angle, and base station height. The input to the base station encoder also includes random noise to enhance the system's diversity and robustness during training and inference.

[0025] Base station encoder, such as Figure 1 As shown, it consists of a base station physical feature input module and a base station parameter decoder.

[0026] The base station physical feature input module is used to input the physical parameters of the base station (such as transmit power, antenna height, downtilt angle, carrier frequency, random noise, etc.) into the system, map or encode them into a latent feature vector, and then input it into the environmental encoder. This feature representation is typically a "high-dimensional vector" required for subsequent network fusion or decoding, designed to capture key information about the base station's impact on wireless signal coverage.

[0027] The base station parameter decoder is based on a diffusion model or other generative framework. It uses random noise and latent vector z to perform a "reverse denoising" or "generation" process, and finally outputs the estimated base station parameters. During training, it compares the parameters with the real base station parameters to generate a loss, thereby constraining the encoder's mapping process.

[0028] The forward diffusion process of the base station parameter decoder includes: gradually adding noise to the data (base station parameters) to move from the initial true values ​​to a Gaussian distribution (or other noise distribution); this can be regarded as a multi-level perturbation of the base station parameters in offline or training mode to simulate uncertainty or measurement error.

[0029] The reverse generation process of the base station parameter decoder includes: a denoising network based on a diffusion model (i.e., the "base station parameter decoder" in this system) starts with pure noise or partially noisy data, iteratively denoises, and finally gradually restores clean base station parameters. This denoising process is mutually constrained or fused with the latent vector z output by the base station physical feature input module to ensure that the restored result matches the true parameters.

[0030] At the base station parameter decoder, random noise is typically introduced in the following ways: Initial noise sampling: During back diffusion, a noise tensor is first sampled from a standard normal distribution, and then base station parameters are generated in reverse through a multi-step denoising network; Fusion encoder features: Several intermediate steps can inject the latent vector z output by the base station physical feature input module into the decoding network, so that the decoder can simultaneously consider the uncertainty of "existing physical prior information z" and "random noise" to generate the final approximate value of base station parameters.

[0031] Reconstruction loss: By comparing the decoder output with the real base station parameters and minimizing the reconstruction error, the base station physical feature input module and the decoder are forced to learn the correct denoising mapping during training and form an accurate restoration of the base station parameters.

[0032] The base station physical feature input module employs a multilayer perceptron (fully connected neural network) structure to embed and nonlinearly transform the original parameters, mapping these numerical parameters into internal feature vector representations. First, the base station parameters are normalized, for example, by scaling values ​​such as transmit power and antenna height to an appropriate range. Then, these normalized parameters are fed as input to several fully connected layers, extracting high-level features layer by layer. The hidden layers of the network can learn the mapping relationship between base station parameters and signal coverage patterns, such as distance attenuation trends and the influence of antenna directivity. In the output layer, the base station physical feature input module generates a low-dimensional feature vector (or feature matrix) that encapsulates the factors influencing signal propagation by the base station.

[0033] During the training process of the base station encoder, after the base station physical feature input module maps the base station physical parameters to the latent vector z, the base station parameter decoder generates a reconstructed base station parameter p from z and random noise through a diffusion model for inverse denoising. If the difference between p and the true parameters is too large, a high reconstruction error (such as mean square error) will occur, which propagates back to the base station physical feature input module, forcing it to learn a more physically meaningful and reversible representation. For the base station parameters of each training sample, noise can be added forward (simulating measurement errors, diverse scenarios), and then the base station parameter decoder can be used to gradually denoise them in reverse. If the denoised result is still close to the original parameters, it means that the model has successfully captured the statistical distribution of the base station parameters and tolerates a certain degree of random perturbation. Base station parameters are often limited by engineering constraints in reality (such as the transmit power being within a certain range, and the downtilt angle not being too large). The diffusion model can reflect these priors in the latent space or the denoising process. If the decoded result exceeds the reasonable domain, a loss penalty will be used to guide the base station physical feature input module-base station parameter decoder to adjust to a range that conforms to physical laws.

[0034] Compared to purely deterministic mappings, diffusion models allow for the preservation of some randomness during the inverse denoising stage, enabling slightly different samples to be generated for the same base station input parameters during training. This simulates minor measurement / input errors in base station deployment, improving model generalization. When training on single-scene or local data, without noise injection and generative mechanisms, the encoder may overfit a specific input-output mapping; with the bidirectional constraint of diffusion denoising, the network is forced to learn the base station parameter distribution more broadly, thus handling the inevitable minor variations in the real environment during inference.

[0035] In actual training, the reconstruction loss generated by the base station parameter decoder often appears together with other loss terms (such as empirical prior loss, adversarial loss, and measured RSRP regression loss).

[0036] During the use of a base station encoder, the base station parameter decoder is used to generate or recover base station parameters. To verify the encoder's fidelity in reproducing base station parameters, or to complete missing / incomplete base station parameters in certain scenarios, the decoder can perform inverse denoising starting from the noise plus z, generating a list of feasible base station parameter candidates. Sometimes, engineering projects require simulating more "virtual base stations" or "base station configurations" to observe coverage changes. The decoder can sample different noise levels, resolving diverse base station parameter instances, which is helpful for large-scale simulation testing.

[0037] The base station parameter decoder is also used for uncertainty quantification: given the same latent vector z, the decoder can sample and denoise noise multiple times during inference to obtain slightly different p, thereby estimating the uncertainty range of the base station parameters. If deviations are found in certain base station parameters (such as antenna height) during actual deployment, the decoder can be updated by retraining or fine-tuning so that the new latent vector reflects this adjustment, thus maintaining the system's accurate representation of the physical environment.

[0038] The base station parameter decoder is primarily used in conjunction with radio map generation: after confirming that the latent vectors of the base station parameters are consistent with the actual parameters, this "physically consistent" z-vector is fused with environmental features and input into the radio map decoder, ultimately generating a more reliable coverage map. When operators change base station configurations (frequency band, power, downtilt angle, etc.), the decoder can also perform denoising under the new parameters, ensuring that the new latent vectors are fused with the environment to output updated coverage predictions, facilitating online or offline planning and evaluation.

[0039] Specifically, such as Figure 2 As shown, the process of generating joint feature representations includes: The environmental image features and feature vectors are fused using an attention module to obtain a fused feature representation; The fused feature representation is then subjected to further feature extraction or refinement, including convolution, upsampling, deconvolution, or multi-scale processing, to obtain a joint feature representation.

[0040] In this embodiment, multiple environmental features from the environmental encoder and base station features from the base station encoder are fused in the attention fusion module to form a joint feature representation. The attention fusion module employs an attention mechanism to combine these two types of features: on the one hand, it uses the feature vector output by the base station encoder as a query or guiding signal to selectively emphasize important parts of the environmental features related to the base station; on the other hand, it modulates the spatial influence of the base station features based on the environmental features (which can be used as keys / values). Through this attention weight allocation, the model can automatically focus on areas in the environmental feature map that play an important role in wireless propagation (e.g., areas related to the line-of-sight distance of the base station location or areas where tall buildings are located).

[0041] During the fusion process, the module can combine convolution, upsampling, and multi-scale feature refinement operations. For example, the fusion module performs upsampling and convolutional fusion on environmental features at different scales, combining local details with global structural information; and uses a multi-scale attention mechanism to simultaneously consider macroscopic path loss trends and microscopic shadow fading details. After processing by the attention fusion module, the output joint feature representation has effectively fused the influence of base station parameters and environmental spatial features, preparing for subsequent decoding to generate a radio power map. This joint feature includes both the overall influence of factors such as base station transmit power, frequency band, and downtilt angle on signal attenuation, and also incorporates the local modulation effects of environmental terrain and buildings on signal distribution.

[0042] In this embodiment, the fused joint features are decoded by a radio map decoder to generate a radio signal power distribution map (e.g., an RSRP intensity plane map) of the target area. The radio map decoder can employ a convolutional decoding network or a generative neural network structure to map the high-dimensional joint features back to a two-dimensional plane corresponding to the geographic area grid. Specifically, the decoder progressively upsamples and deconvolves (transposed convolutions) the joint features, continuously restoring spatial resolution, and finally outputting a power value matrix of the same size as the original environmental distribution map. During the decoding process, the decoder may utilize skip connections to introduce detailed features from earlier levels of the environmental encoder to ensure fine-grained accuracy of the output. The final radio map represents the downlink received signal power intensity (RSRP) at each location within the area in raster form, typically expressed in dBm. Through this decoder, the model reprojects the hidden spatial features into a predicted signal coverage map in the actual geographic space, achieving a transformation from feature space to physical space. The radio power plane map generated in this embodiment can be used to evaluate network coverage effectiveness, identify blind spots and weak areas, etc., and has significant value in network planning and optimization.

[0043] Specifically, the system's training process employs a supervised learning method, combined with prior constraints on physical propagation, to perform end-to-end training. These prior constraints on physical propagation consist of two parts: Empirical prior constraints are used to ensure that the feature vectors output by the base station encoder conform to the theoretical RSRP statistical distribution on a large scale. Deterministic prior constraints are used to constrain the high-frequency details and shadow effects in the target area captured by the radio map output of the environmental encoder and radio map decoder.

[0044] In this embodiment, to fully leverage the complementary advantages of data-driven learning and wireless propagation physics knowledge, the main training process of the model adopts a strategy combining supervised learning and prior constraints on physical propagation, but does not rely on large-scale real RSRP measurement data. The core idea is to use readily available high-precision simulation data (such as coverage distributions generated based on 3D modeling and ray tracing) or radio power distribution maps obtained through professional simulation methods in typical scenarios as the primary "truth value" labels for training.

[0045] To obtain a sufficient number of training samples, this embodiment utilizes high-fidelity simulation tools (such as ray tracing, CDEM scene simulation, etc.) to perform global simulations on several representative regions, generating downlink power distribution maps (RSRP). These simulation data can be considered as approximations of the real situation in the absence of or with very little real measurement data.

[0046] In model training, these simulated data are treated as "ground truth" labels, and the model is trained in a supervised manner. This means that the model is guided to learn by minimizing the difference between the generated radio map and the simulated coverage map (using mean squared error (MSE) or mean absolute error (MAE). This allows for sufficient constraints on the shape, shadow distribution, and other aspects of the radio map at both macroscopic and detailed levels.

[0047] Under this supervised loss, the model learns that for each input (base station parameters + environmental features), the output radio power distribution should closely approximate the power map obtained from ray tracing. This is often the primary form of pre-training because it provides large-scale, high-resolution supervisory signals at the pixel level, helping the model quickly "remember" or "learn" the spatial fading patterns and large-scale / local shading effects simulated by ray tracing.

[0048] Since real RSRP measurement data is not used on a large scale at this time, it will not rely too much on field testing, thus avoiding the problem of "insufficient measurement data leading to unstable training". At the same time, it allows the model to complete the initial learning of wireless propagation laws in a unified and relatively complete simulation environment.

[0049] In addition to performing ordinary pixel-level error measurements (such as MSE / MAE) on the model output and reference (ray tracing results or other simulation data) during training, more refined metrics can be defined for "signal attenuation" and "shadow boundary" and incorporated into the overall loss function, adding constraints for "attenuation magnitude" and "shadow boundary" to the loss function.

[0050] Attenuation amplitude constraint: First, define several “signal attenuation regions” of interest in the ray tracing reference, such as typical areas far from the base station or blocked by buildings; calculate the average attenuation amplitude difference between the model prediction and the reference data in these regions (which can be based on absolute error or relative error); weight this difference as part of the supervision loss, so that the model focuses on the matching of attenuation behavior with the reference in these regions.

[0051] Shadow boundary structure constraints: Using segmentation or edge detection methods, extract the "shadow boundary" region (which can be the edge of a connected region with signal strength below a certain threshold) from the reference map (ray tracing result) and the model output respectively; compare the shape, position or overlap of the two shadow boundary maps, using common edge matching metrics (such as IoU, Dice coefficient, Hausdorff distance, or even SSIM comparison in local structure); introduce the boundary matching degree or similarity into the loss function to guide the model to generate a covered shadow area that is as close as possible to the shape of the ray tracing boundary.

[0052] In practice, the decay magnitude and shadow boundary loss can be added as additional terms and combined with the main supervision loss (such as MSE) into a comprehensive loss. When the network size and data size are sufficient to support it, this multi-constraint fusion can achieve a better balance between details and overall values, and make the model pay explicit attention to "shadow edge" and "decay magnitude".

[0053] The empirical prior, introduced during training, incorporates the signal attenuation curve output by classical propagation models (such as Okumura-Hata, COST-231, etc.) and compares it with the feature vector output by the base station encoder to form an "empirical loss". This loss constraint model maintains consistency with the empirical model in terms of large-scale path loss, avoiding deviation from the basic attenuation law.

[0054] Cell-level radio map generative models attempt to describe the impact of communication base station physical parameters on the generated radio map through base station parameter decoders. Assuming all buildings in the urban area are removed and all objects are replaced with roads, the research scenario degenerates into an ideal open area. In this case, the path fading of radio propagation is only affected by the communication base station physical parameters such as transmit power, center frequency, and base station altitude, as well as the distance between the signal receiving point and the transmitting antenna. This path fading can be described by empirical radio propagation models.

[0055] The introduction of the classical propagation model aims to provide mathematical and statistical constraints for cell-level radio map generative models through empirical physical laws. First, empirical models such as Cost231-Hata or SPM are used to provide the statistical distribution of RSRP estimated by logarithmically fitting the wireless propagation model under specific communication base station parameters. Then, the estimation results of the base station physical feature input module are constrained by a base station parameter decoder based on the diffusion model. Next, the RSRP statistical distribution map of the diffusion model prediction results and the empirical model fitting is calculated using an empirical loss function. By reducing the empirical loss and using the gradient backpropagation algorithm, the RSRP statistical distribution is transformed into guiding information for the base station encoder parameters. In this way, the diffusion model not only helps the model quickly extract useful features of the communication base station physical parameters, but also improves the stability of base station physical feature extraction by relying on prior knowledge of physical laws, further enhancing the adaptability of the generative model in real-world scenarios.

[0056] The deterministic prior: The higher precision ray tracing results obtained through the deterministic ray tracing model can be regarded as deterministic priors.

[0057] The environmental encoder and radio map decoder form the core modules of the cell-level radio map generative model. Radio propagation in urban areas is subject to environmental obstruction, resulting in significant reflections and diffraction. These radio multipath and shadowing effects can be accurately captured by a deterministic ray tracing model. The deterministic ray tracing model can provide rich high-frequency details from the radio map, thus guiding the environmental encoder and radio map decoder to generate detailed radio maps. Therefore, radio maps calculated by the deterministic ray tracing model are added as auxiliary labels to the model training set.

[0058] The specific approach involves generating adversarial training. The discriminator treats the radio map generated by ray tracing as a "real" sample, guiding the generator's (this model's) output to continuously approach the physically simulated values, thereby capturing fine propagation effects such as building shadows and multipath propagation. Adversarial training helps the model learn global distribution characteristics and local fading textures that more closely approximate real propagation phenomena, supplementing and enhancing the simple pixel-level errors of supervised training.

[0059] In the overall training objective function, the supervised loss (simulated data true value), empirical prior loss, and adversarial loss can be combined to form a multi-objective weighted comprehensive loss function.

[0060] The training process employs iterative optimization. On the one hand, it minimizes the difference between the model output and the simulated true value, allowing the model to maintain a global approximation of the simulation results. On the other hand, it uses prior loss constraints to promote the model to simultaneously follow empirical and high-precision physical propagation laws, thereby enhancing its generalization and physical rationality.

[0061] Through the above strategies, the model can efficiently learn the wireless propagation patterns in a wide range of scenarios during the "main training" phase by utilizing high-fidelity simulated coverage maps and physical priors, overcoming the limitation of insufficient real RSRP measurement data.

[0062] Specifically, the empirical prior constraints are implemented through the following steps: The theoretical RSRP statistical distribution is generated using the physical parameters of the communication base station and an empirical logarithmic fitting model. Based on this, an empirical loss is constructed. This empirical loss directly constrains the feature vector output by the base station encoder, making it consistent with the theoretical RSRP statistical distribution in terms of large-scale path attenuation.

[0063] Empirical priors are predictions of signal strength based on existing classical wireless propagation empirical models (such as the free space path loss model, the Okumura-Hata model, or the COST-231 model), and these predictions are introduced as constraints on the model.

[0064] The specific steps are as follows: Based on the base station's transmission parameters and basic environmental characteristics, a theoretical RSRP distribution plane is calculated using a selected empirical propagation model. This theoretical distribution reflects the approximate trend of received signal strength attenuation with distance under typical empirical conditions, without considering complex multipath effects.

[0065] Subsequently, the representation of the base station encoder output within the model is compared with the theoretical RSRP distribution generated by the empirical model, and a corresponding loss function is defined (e.g., calculating the mean square error of the difference between the two). Since the output of the base station encoder implicitly contains the coverage contribution of the base station to the entire area, introducing this loss term is equivalent to adding a physical constraint correction to the output of the base station encoder: forcing the signal distribution it represents to conform to the attenuation behavior expected by the empirical formula.

[0066] Minimizing this loss during training prompts the model to adjust base station-related features so that its initial signal coverage pattern closely approximates the predictions of the empirical model. This constraint ensures that the model can still follow basic path loss rules even when sufficient data is lacking, avoiding predictions that contradict known physical facts. By incorporating empirical priors, this embodiment significantly improves the model's ability to capture macroscopic signal power distribution, making the generated radio map more reasonable and reliable in its overall trend.

[0067] Specifically, the deterministic prior constraints are implemented through the following steps: An auxiliary radio map is generated using a deterministic ray tracing model, which reflects high-frequency details caused by building occlusion, multipath, and shadowing effects within the target area; An environment encoder and a radio map decoder are combined to form a generator network, and a discriminator network is configured. During generative adversarial training, the auxiliary radio map is used as a real sample and compared with the output of the generator network to impose adversarial constraints on the outputs of the environment encoder and the radio map decoder.

[0068] Deterministic priors leverage high-precision ray tracing simulation results to enhance the model's characterization of complex propagation phenomena and are integrated into model learning through an adversarial training mechanism. The deterministic priors originate from ray-traced-assisted generated radio maps, which are approximately realistic RSRP distribution maps obtained through ray tracing technology given a 3D environmental model and base station parameters. Ray tracing considers deterministic physical effects in electromagnetic wave propagation (such as reflection, refraction, and diffraction), thus the generated signal distribution is considered a "real" sample. To further approximate the ray tracing effect in the model output, this embodiment employs the Generative Adversarial Network (GAN) approach, using the ray-traced distribution as a real sample for the discriminator and introducing adversarial loss to constrain model training.

[0069] Specifically, a discriminator network is added during training. This discriminator takes a radio power distribution map as input and determines whether its source is model-generated (predicted) or ray-tracing simulation (real). The model's radio map decoder acts as a generator, producing a predicted power distribution map and attempting to fool the discriminator. Through alternating training, the discriminator improves its ability to distinguish between real and fake distributions, while the generator continuously adjusts its output by minimizing adversarial loss, making its distribution characteristics closer to the real distribution of ray tracing and harder for the discriminator to distinguish.

[0070] This adversarial training method ensures that the model's detailed depiction of signal distribution is closer to physical reality: for example, complex phenomena such as rapid signal attenuation in the shadow areas of tall buildings and slow signal changes in open areas can be more accurately reflected in the model output. By combining adversarial training with deterministic priors, this embodiment significantly improves the credibility and realism of the generated radio maps, enabling the model to not only capture the overall trend but also depict detailed changes.

[0071] Example 2 This invention provides a method for generating a two-dimensional radio map that integrates communication base station and environmental information of the system, comprising the following steps: The environment encoder converts the environment map of the target area into environment image features; The physical parameters of the communication base stations in the target area are converted into feature vectors using a base station encoder. Within the environmental encoder or in subsequent fusion layers, the environmental image features are fused with the feature vector output by the base station encoder to generate a joint feature representation. A radio signal power plane distribution map of the target area is generated by a radio map decoder based on the joint feature representation.

[0072] Specifically, the method steps of the system described in this invention include the following processes: Environmental Feature Extraction: Acquire environmental distribution map data of the target area (including at least three rasterized distribution maps: land cover type, building height, and ground elevation), input it into the environmental encoder, and extract multi-level environmental image feature representations. The features output by the environmental encoder retain information that affects signal propagation, such as topographic relief and building distribution within the area.

[0073] Base station feature extraction: Relevant base station parameters (carrier frequency, antenna transmit power, vertical downtilt angle, base station height, etc.) and an additional random noise vector are input into the base station encoder to extract the base station parameter feature vector. This feature vector condenses the characteristics of the base station's transmitted signal and includes random perturbations to enhance robustness.

[0074] Feature fusion processing: The environmental features output by the environmental encoder and the base station features output by the base station encoder are fed into the attention fusion module for fusion. An attention mechanism is used to weight the environmental features, guiding the focus on key spatial locations through base station features. Combined with multi-scale feature refinement operations such as convolution and upsampling, a fused joint feature representation is generated. This joint feature fully integrates information from both the environment and the base station.

[0075] Radio map generation: The above-mentioned joint features are input into the radio map decoder. After upsampling and decoding operations, a radio signal power planar map (RSRP distribution map) of the target area is generated. This power distribution map gives the predicted signal power at each location within the area in the form of a two-dimensional grid.

[0076] This embodiment can be applied to cell-level wireless network planning and optimization tasks in typical complex urban environments. For example, in urban cellular network deployments with dense high-rise buildings and crisscrossing streets, this embodiment can be used to precisely model the wireless coverage range of each cell, fully considering the impact of terrain obstruction and signal overlap interference between neighboring cells. Through a data-driven high-precision propagation model and automatic optimization algorithm, the system can perform parallel optimization for hundreds or thousands of cell sites, adjusting parameters such as base station power and antenna downtilt angle in a short time to eliminate weak coverage areas.

[0077] Compared to traditional planning methods that rely on experience and extensive drive testing, this embodiment improves the accuracy of coverage prediction and reduces error accumulation caused by model simplification. For example, high-fidelity simulation modeling can significantly reduce the cost of pilot site construction during the planning phase, while reducing on-site testing workload by more than 30%. Automated cell-level optimization significantly improves network deployment efficiency and significantly reduces operation and maintenance costs.

[0078] Therefore, this embodiment has engineering deployment feasibility and outstanding practical value in urban cellular network optimization, and can complete network coverage optimization in complex scenarios at a lower cost and higher efficiency while ensuring user experience.

[0079] Example 3 The present invention provides a computer storage medium, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the above technical solution.

[0080] Example 4 The present invention also provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described in the above technical solution.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0086] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A two-dimensional radio map generation system integrating communication base station and environmental information, characterized in that: include: A base station encoder is used to convert the physical parameters of communication base stations within a target area into feature vectors that reflect the characteristics of electromagnetic relay. An environment encoder is used to convert a map of a target area containing environmental information into environmental image features that reflect the global characteristics of the target area; and to fuse the environmental image features and feature vectors to generate a joint feature representation. A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation.

2. The system according to claim 1, characterized in that: The map of the target area containing environmental information includes a map of the distribution of land features, a map of building heights, and a map of ground elevation.

3. The system according to claim 1, characterized in that: The process of generating joint feature representations includes: The environmental image features and feature vectors are fused using an attention module to obtain a fused feature representation; The fused feature representation is then subjected to further feature extraction or refinement, including convolution, upsampling, deconvolution, or multi-scale processing, to obtain a joint feature representation.

4. The system according to claim 2, characterized in that: The physical parameters of a communication base station include its location, carrier frequency, antenna transmit power, vertical downtilt angle, and / or base station height.

5. The system according to claim 1, characterized in that: The input to the base station encoder also includes random noise to enhance the system's diversity and robustness during training and inference.

6. The system according to claim 1, characterized in that, The system's training process employs a supervised learning method, combined with prior constraints on physical propagation, to perform end-to-end training. These prior constraints on physical propagation consist of two parts: Empirical prior constraints are used to ensure that the feature vectors output by the base station encoder conform to the theoretical RSRP statistical distribution on a large scale. Deterministic prior constraints are used to constrain the high-frequency details and shadow effects in the target area captured by the radio map output of the environmental encoder and radio map decoder.

7. The system according to claim 6, characterized in that, The empirical prior constraints are implemented through the following steps: The theoretical RSRP statistical distribution is generated using the physical parameters of the communication base station and an empirical logarithmic fitting model. Based on this, an empirical loss is constructed. This empirical loss directly constrains the feature vector output by the base station encoder, making it consistent with the theoretical RSRP statistical distribution in terms of large-scale path attenuation.

8. The system according to claim 6, characterized in that, The deterministic prior constraints are implemented through the following steps: An auxiliary radio map is generated using a deterministic ray tracing model, which reflects high-frequency details caused by building occlusion, multipath, and shadowing effects within the target area; An environment encoder and a radio map decoder are combined to form a generator network, and a discriminator network is configured. During generative adversarial training, the auxiliary radio map is used as a real sample and compared with the output of the generator network to impose adversarial constraints on the outputs of the environment encoder and the radio map decoder.

9. A method for generating a two-dimensional radio map based on the system of any one of claims 1-8, integrating communication base station and environmental information, characterized in that, Includes the following steps: The environment encoder converts the environment map of the target area into environment image features; The physical parameters of the communication base stations in the target area are converted into feature vectors using a base station encoder. Within the environmental encoder or in subsequent fusion layers, the environmental image features are fused with the feature vector output by the base station encoder to generate a joint feature representation. A radio signal power plane distribution map of the target area is generated by a radio map decoder based on the joint feature representation.

10. The present invention also provides an electronic device, comprising: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, characterized in that the memory stores computer instructions, and the processor executes the computer instructions to perform the method of claim 9.

Images