A three-dimensional radio map construction method based on conditional generative adversarial network

Abstract

This invention discloses a method for constructing a 3D radio map based on a conditional generative adversarial network (GAN), comprising the following steps: Step S1: Acquire multimodal input data of the target area; Step S2: Based on generalized anisotropic spatial attenuation field coding technology, map discrete transmitter points to continuous spatial energy distribution features, generate a global radiation prior feature map, and construct a conditional feature tensor; Step S3: Construct a 3D U-Net generator with a self-attention mechanism and establish a GAN loss function based on composite constraints; Step S4: Iteratively train the generator and discriminator using the constructed training dataset; Step S5: Online inference and real-time construction of a high-precision 3D radio map. This invention can integrate complex 3D geographic environment information and real-time multimodal data to achieve 3D radio map construction with high accuracy, deterministic mapping, and low latency.

Description

Technical Field

[0001] This invention relates to the field of radio technology, and in particular to a method for constructing three-dimensional radio maps based on conditional generative adversarial networks. Background Technology

[0002] In the era of 6G communication and low-altitude networks, real-time monitoring and management of radio resources are crucial for supporting diverse intelligent applications, such as drone swarm trajectory planning and remote collaborative robots. Radio map estimation (RME) is a core technology for characterizing the distribution of radio resources. In recent years, deep learning (DL)-based methods have been widely applied in the RME field due to their powerful capabilities in feature representation learning and efficient inference speed, significantly outperforming traditional ray tracing and data interpolation methods.

[0003] However, most current DL-based RME methods are limited to two-dimensional (2D) near-Earth scenarios and fail to capture the essential three-dimensional (3D) signal propagation characteristics (such as channel fading and antenna polarization effects along the height dimension), which limits their application in low-altitude 3D networks. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a three-dimensional radio map based on a conditional generative adversarial network. This method can integrate complex 3D geographic environment information and real-time multimodal data to achieve 3D radio map construction with high accuracy, deterministic mapping and low latency.

[0005] The objective of this invention is achieved through the following technical solution: a method for constructing a three-dimensional radio map based on conditional generative adversarial networks, comprising the following steps:

[0006] Step S1: Acquire multimodal input data for the target region;

[0007] Step S2: Based on the generalized anisotropic spatial attenuation field coding technique, the discrete transmitter points are mapped to continuous spatial energy distribution features, generating a global radiation prior feature map and constructing a conditional feature tensor.

[0008] Step S3: Construct a 3D U-Net generator with a self-attention mechanism and establish a generative adversarial network loss function based on composite constraints;

[0009] Step S4: Use the constructed training dataset to iteratively train the generator and discriminator alternately;

[0010] Step S5: Online reasoning and real-time construction of high-precision 3D radio maps.

[0011] The beneficial effects of this invention are: 1. Extremely fast single-step deterministic inference (overcoming the problems of high latency and strong uncertainty in diffusion models): Due to the adoption of an end-to-end 3D U-Net generator structure, this invention does not require multi-step iterative denoising as in diffusion models during the inference phase. The network can deterministically output a high-fidelity 3D radio map in milliseconds through a single forward propagation (One-shot inference), greatly reducing computational complexity and perfectly adapting to the extremely low latency requirements of real-time radio resource management in 6G and low-altitude networks.

[0012] 2. High-precision 3D spatial feature perception and representation (overcoming the lack of spatial awareness in 2D models): By designing a 3D convolutional kernel in the encoder of 3DU-Net that "preserves height dimensional resolution," the model can accurately capture the vertical fading features of signals as altitude changes. Simultaneously, the introduced bottleneck layer self-attention mechanism enables the model to overcome the limitations of local convolution and successfully extract long-distance channel dependencies in 3D space. The combination of these two features results in estimation accuracy far exceeding that of existing 2D models in 3D urban environments (including complex building occlusion and multipath effects).

[0013] 3. Effectively solves training collapse caused by extremely sparse input (overcoming feature sparsity problem): This solution innovatively converts discrete, extremely sparse transmitter coordinates into a 3D Gaussian heatmap. This technique not only effectively eliminates the gradient vanishing problem that is easily caused by sparse features in continuous convolution and pooling operations in deep networks, but also enables the neural network to better learn the spatial correlation between the transmitter's physical location and the signal propagation and radiation pattern, significantly improving the model's convergence speed and estimation robustness.

[0014] 4. High Preservation of Local Signal Details and Physical Consistency: A patch-level 3D CNN discriminator is used to evaluate the physical plausibility of local regions, and a voxel-level L1 loss function is combined to constrain the generator. This adversarial training mechanism effectively avoids the common problem of blurred boundaries in generated images, ensuring high-quality preservation of local details in high-contrast areas such as building edges and signal shadows in the generated 3D radio map. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0017] This invention specifically designs a 3D U-Net with a self-attention mechanism as a generator for 3D radio map estimation tasks, which can perfectly capture the fading characteristics and spatial dependencies of signals in three-dimensional space (especially the height dimension). Secondly, to address the network training instability and gradient vanishing problems caused by the extreme sparseness of transmitter location information in traditional methods, we propose a novel 3D Gaussian heatmap encoding technique to reconstruct the input features. Finally, we introduce a patch-level 3D convolutional neural network as a discriminator and conduct co-adversarial training by introducing voxel-level L1 loss and dynamically scheduled adversarial loss. After training, the generator can achieve high-precision, deterministic, and ultra-fast one-shot inference from multimodal inputs (environment, transmitter, sparse measurements) to a 3D radio map.

[0018] Data preprocessing: The input 3D geographic environment map, transmitter information, and sparse measurements are transformed into multi-channel 3D image tensors. For the extremely sparse transmitter location information, the traditional multi-hot representation is abandoned, and it is encoded as a 3D Gaussian heatmap centered on the transmitter coordinates (based on a ternary Gaussian distribution). This eliminates the vanishing problem of sparse features in continuous convolutional pooling, helping the model better learn the physical correlation between the transmitter and signal propagation patterns.

[0019] Generator Design: An improved 3D U-Net model is used as the generator. The 3D convolutional layers in its encoder are specially designed to avoid downsampling in the height dimension, so as to preserve the vertical high-resolution information of the signal as the altitude changes; a self-attention module is introduced at the bottleneck of the network to capture long-range feature dependencies in 3D space; the decoder fuses high-resolution local information through skip connections to output a high-fidelity 3D radio map image.

[0020] Discriminator Design: A conditional network model based on multi-layer 3D convolution is used as the discriminator. Its input includes not only generated or real 3D radio maps, but also conditional features (geographic maps, Gaussian heatmaps, etc.). The discriminator performs patch-level discrimination at the end of the network, relying not only on global statistical data, but also on finer evaluation of the structural consistency and physical plausibility of local signals in 3D space.

[0021] GAN learning algorithm: Guides network adversarial training by constructing a specific loss function. The generator's loss function consists of a dynamically weighted adversarial loss and a voxel-level L1 loss (L1 Loss), where the L1 loss effectively preserves the boundary features of local signals in the radio map. The discriminator's loss function includes adversarial loss and gradient penalty regularization loss. During training, a staged scheduling strategy is used for the weights of the adversarial loss to prevent the discriminator from prematurely suppressing the generator, ensuring training stability.

[0022] Model Training and Inference: During the training phase, the generator and discriminator perform joint adversarial optimization based on the constructed 3D dataset. The generator continuously generates higher-quality radio maps to "deceive" the discriminator, and the discriminator continuously improves its discrimination ability. During the inference phase, the discriminator is removed, and only the trained 3D U-Net generator is used to achieve fast end-to-end forward inference and output the estimation results of the 3D radio map in real time.

[0023] Specifically, such as Figure 1 As shown, a method for constructing a 3D radio map based on a conditional generative adversarial network includes the following steps:

[0024] Step S1: Acquire multimodal input data for the target region;

[0025] In complex 3D communication scenarios, the three-dimensional geographic environment (discrete obstacle mesh), the received signal strength (extremely sparse single-point scalar), and the transmitter information (spatial coordinates and power parameters) are heterogeneous multimodal data. Traditional deep learning networks have difficulty aligning and jointly extracting features from these data with different feature domains and extremely different sparsity.

[0026] 3D geographic environment information: Acquire a 3D map of the target area, convert it into a voxel mesh with uniform spatial resolution, and generate a 3D binary occupancy tensor. Where 1 indicates that the voxel is occupied by an obstacle, and 0 indicates free space; W, D, and H represent the width, depth, and height of the target area, respectively.

[0027] Sparse measurement tensor Within a voxel grid with uniform spatial resolution in the target region, at a preset sampling rate... Perform uniform sampling to obtain a set of sampling locations. Elements at unsampled voxel grid locations are set to 0, while elements at sampled voxel grid locations are set to the actual received signal strength, thus constructing a sparse measurement tensor. ;

[0028] Transmitter Information , obtain A set of three-dimensional coordinates of a transmitter and its transmission power Construct the transmitter information tensor If a transmitter exists at a location in the voxel mesh, then the element value at that location is... Otherwise, it is 0.

[0029] Step S2: Based on the generalized anisotropic spatial attenuation field coding technique, the discrete transmitter points are mapped to continuous spatial energy distribution features, generating a global radiation prior feature map and constructing a conditional feature tensor.

[0030] Traditional radio map estimation methods directly input the discrete coordinates of the transmitter, which exhibits an "extremely sparse" characteristic in three-dimensional space. In the continuous convolution and downsampling pooling operations of deep networks, such single-point features are prone to "gradient vanishing," causing the model to fail to establish the physical correlation between the transmitter source and the signal propagation across the entire spatial domain.

[0031] Encoding is performed using a generalized anisotropic spatial attenuation field (SAAField) coding technique. For each transmitter... Instead of using point coordinates, it uses their physical coordinates. Establish a continuous three-dimensional attenuation feature map centered on the spatial volume. The eigenvalue at point is calculated as follows:

[0032] After the calculation is completed, the attenuation feature maps generated by all transmitters are normalized and aggregated to generate a global radiation prior feature map. .

[0033] Through parameters Controlling the decay pattern and using anisotropic weight vectors Different attenuation rates are assigned to different dimensions of space. This mechanism cleverly transforms a discrete signal source into a continuous energy prior distribution that conforms to the physical laws of electromagnetic wave spatial diffusion. Anisotropic weights are introduced. It can perfectly simulate the fading difference of signals propagating horizontally and vertically in the real world (such as the vertical polarization effect of antennas). This fundamentally overcomes the problem of vanishing sparse features and provides powerful physical space prior guidance for neural networks.

[0034] Transform the three-dimensional binary occupancy tensor sparse measurement tensor Normalization is performed across the entire dataset, mapping it to a unified physical coordinate system, and converting it into grayscale image pairs containing multiple channels, which are then compared with the global radiative prior feature map. The final concatenation constructs a conditional feature tensor. .

[0035] By constructing a unified high-dimensional physical coordinate tensor, the spatial mismatch problem between different modalities of data is eliminated. This step provides a data foundation with strict physical correspondence for subsequent 3D fully convolutional networks, enabling the network to simultaneously perceive both "obstacle occlusion" and "signal strength information" at the same spatial location, significantly reducing the training threshold for heterogeneous data fusion.

[0036] Step S3: Construct a 3D U-Net generator with a self-attention mechanism and establish a generative adversarial network loss function based on composite constraints;

[0037] When performing feature downsampling, conventional 3D convolutional neural networks (3D CNNs) often compress features equally in all three spatial dimensions. This leads to the loss of resolution information in the "vertical height dimension (Z-axis)" that is crucial for low-altitude network applications. At the same time, traditional convolutions are limited by local receptive fields and cannot capture long-distance multipath effects and complex building occlusion (Macro-shadowing) dependencies that span an entire block.

[0038] Therefore, an improved 3D U-Net generator with high-fidelity constraints and a self-attention mechanism is designed.

[0039] (1) The encoder contains multiple downsampling layers, each downsampling layer includes a downsampling module, and the core operation of the downsampling module is represented as follows:

[0040] ;

[0041] By configuring asymmetric convolution kernels, the height dimension can be forcibly removed. Downsampling operation on: The convolutional kernel is designed not to be in the height dimension. Downsampling is performed on the signal to preserve the vertical distribution information of the signal as altitude changes; Indicates the layer normalization function; For the activation function; input After performing downsampling operations layer by layer through multiple downsampling layers, the feature map output by each downsampling layer is obtained;

[0042] The feature map output from the last downsampling layer is passed to the depth bottleneck layer, which then outputs the feature map. and in the feature map Introducing a self-attention module to capture long-distance dependencies:

[0043] Calculate the projection matrix:

[0044]

[0045] Represents query, key, and value; , , Weights for queries, keys, and values

[0046] Attention feature output is calculated as follows :

[0047]

[0048] Finally, features are fused using residual connections: ,in For learnable scaling parameters;

[0049] (2) The decoder contains multiple upsampling layers and restores spatial resolution through transposed convolution:

[0050] Will The input is fed into the decoder, where it is upsampled layer by layer through multiple upsampling layers to obtain the feature map output by each upsampling layer;

[0051] By using skip connections, the feature maps output from each upsampling layer in the decoder are concatenated with the feature maps output from the corresponding downsampling layers in the encoder. All concatenated results are then fused using 3D convolutional layers to finally output the predicted 3D radio map image. ;

[0052] Due to the downsampling module in the encoder In the middle, configure asymmetric convolution kernels to forcibly remove the height dimension. Downsampling operation. Specifically, The convolutional kernels are specifically designed not to be in the height dimension. Downsampling is performed to preserve the vertical distribution information of the signal as it changes with altitude; secondly, the output feature map of the network's deep bottleneck layer is... We introduce a self-attention module to capture long-distance dependencies.

[0053] The asymmetric convolutional design, which preserves the height dimension resolution, ensures that the model can accurately reproduce the vertical gradient variation of the signal along the altitude at any network depth. Meanwhile, the self-attention mechanism, by calculating the correlation matrix of global pixels, breaks the local physical limitations of traditional convolutional kernels, enabling the model to perceive the shadow occlusion caused by distant obstacles at the current position. Skip connections are responsible for re-injecting the preserved original high-frequency spatial structure into the decoding process, achieving high-fidelity map generation.

[0054] Considering that a single voxel-level regression loss (such as simple L1 or MSE) can cause a severe "mean regression" problem when generating images, resulting in blurred building shadow edges in radio maps; and that it is extremely sensitive to random noise outliers in real measurements. In addition, conventional adversarial losses lack fine-grained constraints on local physical consistency.

[0055] A multi-dimensional and composite adversarial loss evaluation system is constructed. The generator adopts... It integrates smooth L1 loss (Huber Loss) and 3D structural similarity loss (3D-SSIM). Provides a stable linear gradient when the prediction error is large, avoiding gradient explosion; 3D SSIM loss. This ensures the generator maintains its physical structural characteristics at building edges and signal coverage boundaries. The specific definitions of each component are as follows:

[0056]

[0057]

[0058] in Represents voxel residuals, For the threshold; Realistic 3D radio map image; Represents the structural similarity loss function;

[0059] Discriminator loss function Includes standard adversarial loss and regularization loss to distinguish between real and fake samples:

[0060]

[0061] Regularization loss Used to improve training stability These are the model parameters for the discriminator. As expected, is the hyperparameter for regularization loss.

[0062] It provides a smooth gradient when the error is small, and switches to linear penalty when the error is large (encountering abnormal noise), thus solving the model robustness problem; The forced generator focuses on the brightness and structural similarity of adjacent pixels, enabling the model to generate boundaries that conform to the laws of physical refraction / diffraction at signal abrupt changes such as building edges.

[0063] Step S4: Iteratively train the generator and discriminator using the constructed training dataset; the constructed training dataset contains several training samples, each training sample containing a conditional feature tensor. Its corresponding real 3D radio map image ;

[0064] 1. Fix the generator, calculate the discriminator loss function, and update the discriminator network parameters using gradient descent.

[0065] 2. Fix the discriminator, calculate the generator loss function, and update the generator network parameters using gradient descent.

[0066] 3. Phased Scheduling Strategy: In the early stages of training (e.g., the first 15 epochs), the weight of the adversarial loss in the generator loss function is set to 0, allowing the generator to learn the basic signal distribution first through L1 loss. Subsequently, over a certain period (e.g., 10 epochs), the weight of the adversarial loss is linearly increased to a target value (e.g., 0.1) and maintained. This strategy prevents the discriminator from converging too quickly in the early stages, thus suppressing the generator's learning. The above training process is repeated until the network converges.

[0067] Step S5: Online reasoning and real-time construction of high-precision 3D radio maps.

[0068] In practical wireless network resource management applications, the discriminator is removed, and only the deployed, trained 3DU-Net generator is loaded onto the edge server or cloud computing platform. Multimodal input data of the current target area is acquired in real time, and the conditional feature tensor, after preprocessing in steps S1 and S2, is input to the generator. The generator, through a single-shot inference, can output a high-precision 3D radio map of the area within milliseconds.

[0069] In the embodiments of this application, the multimodal input features can be flexibly replaced: in addition to simultaneously inputting "transmitter information + sparse measurement values ​​+ environmental map", the system can downgrade the input to "sparse measurement values ​​only + environmental map" or "transmitter information only + environmental map" depending on the availability of actual scene data. The generator structure remains unchanged, only the number of input channels is changed, and a 3D radio map can still be output.

[0070] For spatial diffusion coding of transmitter information, in addition to the ternary Gaussian distribution described in this scheme, other forms of distance decay functions (such as exponential decay function, inverse polynomial decay function, etc.) can be used, as long as they can convert single-point sparse information into a continuous spatial gradient distribution.

[0071] In some applications where smoothness requirements are higher, the voxel-level L1 loss in the generator loss function can be replaced by mean squared error loss (L2 loss / MSE) or smoothed L1 loss (Huber loss).

[0072] The standard self-attention module in the bottleneck layer of 3D U-Net can be replaced by a linear attention mechanism or a window-based local attention mechanism (such as the Swin-Transformer Block) with lower computational complexity to further reduce the memory usage of large-scale 3D maps.

[0073] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1.A method for constructing a three-dimensional radio map based on a conditional generative adversarial network, characterized in that: Includes the following steps: Step S1: Acquire multimodal input data for the target region; Step S2: Based on the generalized anisotropic spatial attenuation field coding technique, the discrete transmitter points are mapped to continuous spatial energy distribution features, generating a global radiation prior feature map and constructing a conditional feature tensor. S201. Map transmitter information to a unified physical space coordinate system, then: For each transmitter With its physical coordinates A continuous three-dimensional attenuation feature map is established around the center. In the three-dimensional attenuation feature map, the spatial volume... The eigenvalue at point is calculated as follows: ； an exponential hyperparameter for controlling the decay of the tail shape, an anisotropy weight vector for controlling the decay rates in the three spatial dimensions; After the calculation, the attenuation feature maps generated by all the transmitters are normalized and spliced to generate a global radiation prior feature map ; S202. Convert the three-dimensional binary occupancy tensor sparse measurement tensor Normalization is performed across the entire dataset, mapping it to a unified physical coordinate system, and converting it into grayscale image pairs containing multiple channels, which are then compared with the global radiative prior feature map. The final concatenation constructs a conditional feature tensor. ; Step S3: Construct a 3D U-Net generator with a self-attention mechanism and establish a generative adversarial network loss function based on composite constraints; Step S4: Use the constructed training dataset to iteratively train the generator and discriminator alternately; Step S5: Online reasoning and real-time construction of high-precision 3D radio maps. 2.The method of claim 1, wherein the method comprises: The step S1 of acquiring multimodal input data of the target region includes: 3D geographic environment information: Acquire a 3D map of the target area, convert it into a voxel mesh with uniform spatial resolution, and generate a 3D binary occupancy tensor. Where 1 indicates that the voxel is occupied by an obstacle, and 0 indicates free space; W, D, and H represent the width, depth, and height of the target area, respectively. Sparse measurement tensor Within a voxel grid with uniform spatial resolution in the target region, at a preset sampling rate... Perform uniform sampling to obtain the set of sampling locations. Elements at unsampled voxel grid locations are set to 0, while elements at sampled voxel grid locations are set to the actual received signal strength, thus constructing a sparse measurement tensor. ; Transmitter Information , obtain A set of three-dimensional coordinates of a transmitter and its transmission power Construct the transmitter information tensor If a transmitter exists at a location in the voxel mesh, then the element value at that location is... Otherwise, it is 0. 3.The method of claim 1, wherein: Step S3 includes: S301. Construct a 3D U-Net generator, including: (1) The encoder contains multiple downsampling layers, each downsampling layer includes a downsampling module, and the core operation of the downsampling module is represented as follows: ； By configuring asymmetric convolution kernels, the height dimension can be forcibly removed. Downsampling operation on: The convolutional kernel is designed not to be in the height dimension. Downsampling is performed on the signal to preserve the vertical distribution information of the signal as altitude changes; Indicates the layer normalization function; For the activation function; input That is, the conditional feature tensor is downsampled layer by layer through multiple downsampling layers to obtain the feature map output by each downsampling layer; The feature map output from the last downsampling layer is passed to the depth bottleneck layer, which then outputs the feature map. and in the feature map Introducing a self-attention module to capture long-distance dependencies: Calculate the projection matrix: ； representing a query, key, value; , , weight for query, key, value The attention feature output is calculated as : ； Finally, the features are fused through a residual connection: where are learnable scaling parameters; (2) The decoder contains multiple upsampling layers and restores spatial resolution through transposed convolution: Will The input is fed into the decoder, where it is upsampled layer by layer through multiple upsampling layers to obtain the feature map output by each upsampling layer; By using skip connections, the feature maps output from each upsampling layer in the decoder are concatenated with the feature maps output from the corresponding downsampling layers in the encoder. All concatenated results are then fused using 3D convolutional layers to finally output the predicted 3D radio map image. ; S302. Establish the loss function for generative adversarial networks based on composite constraints: Design loss functions for the generator and discriminator respectively to guide adversarial training: Generator loss function The generator loss function is the adversarial loss fed back from the discriminator. Smoothing L1 loss and 3D structural similarity loss The weighted average is calculated using the following formula: ； wherein, Provide stable linear gradient when prediction error is large, avoid gradient explosion; three-dimensional space SSIM loss Make the generator keep the physical structure features at the building edge and signal coverage junction. The specific definitions of each component are as follows: ； ； wherein represents a voxel residual, is a threshold value; a real three-dimensional radio map image; denotes a structural similarity loss function; discriminator loss function : contains the standard adversarial loss that distinguishes between real and fake samples and a regularization loss ； regularization loss , for improving training stability, are model parameters of the discriminator, is an expectation, is a hyperparameter of the regularization loss. 4.The method of claim 3, wherein: In the step S4, the constructed training data set contains a plurality of training samples, each of which contains a conditional feature tensor a real three-dimensional radio map image corresponding thereto The training process is as follows: S401. Fix the generator, calculate the discriminator loss function, and update the discriminator network parameters using gradient descent. S402. Fix the discriminator, calculate the generator loss function, and update the generator network parameters using gradient descent. S403. A phased scheduling strategy is adopted: During the initial training period, the weight of the adversarial loss in the generator loss function is set to 0, so that the generator learns the basic signal distribution through L1 loss first. Subsequently, the weights of the adversarial loss are linearly increased to the target value and maintained within a set period to prevent the discriminator from converging too quickly in the early stages, thereby suppressing the learning of the generator. Repeat the above training process until the network converges. 5.The method of claim 1, wherein the method is based on a conditional generative adversarial network. Step S5 includes: Remove the discriminator and load the deployed, trained 3D U-Net generator onto an edge server or cloud computing platform; The multimodal input data of the current target region is acquired in real time, and the conditional feature tensor obtained after preprocessing in steps S1 and S2 is input to the generator; The generator outputs a high-precision 3D radio map of the area through a single forward propagation.

Images