A three-dimensional spectrum situation completion method for complex urban electromagnetic environment

Abstract

The application discloses a three-dimensional spectrum situation completion method for a complex urban electromagnetic environment. The method generates a three-dimensional building mask according to building distribution, generates a two-dimensional sampling mask in a non-building area and broadcasts along the height dimension; feature extraction and intra-layer fusion are performed on the sparse spectrum situation, the sampling mask and the building mask layer by layer along the height, and then the fusion features of each height layer are spliced along the channel dimension and are subjected to cross-height layer convolution fusion to obtain conditional features; during training, the conditional features are added to the noisy features of the complete spectrum situation and then input into a U-Net noise prediction network, and the added noise is used as a supervision target; during inference, starting from random noise, cross-step reverse denoising is performed based on the conditional features corresponding to actual sparse monitoring data, and a complete three-dimensional spectrum situation is output, which is used for constructing a three-dimensional radio environment map.

Description

Technical Field

[0001] This invention relates to the field of radio map generation technology, and in particular to a three-dimensional spectrum situational awareness method for complex urban electromagnetic environments. Background Technology

[0002] With the rapid development of the wireless communication industry, electronic devices such as radios and radars have created a dynamic, complex, and multi-dimensional electromagnetic environment. The diversified wireless data transmission services have led to an increase in frequency-using devices and a surge in the probability of frequency conflicts. To rationally allocate spectrum resources and ensure that the service and control signals of each frequency-using unit obtain secure and reliable wireless communication links during transmission, it is necessary to construct a Radio Environment Map (REM) that reflects the distribution of received signal strength in different spatial dimensions, thereby achieving visualization and refined management of the electromagnetic environment.

[0003] Existing REM (Resonance Mode) construction methods mainly fall into two categories: model-driven and data-driven. Model-driven methods rely on electromagnetic propagation theory, using prior knowledge to infer the spectral state distribution of the target area. However, it is difficult to create a universally applicable signal propagation model, and they are easily affected by non-line-of-sight propagation in complex backgrounds such as urban areas. Data-driven methods use spatial interpolation or machine learning techniques to reconstruct a complete signal map from a sparse measurement set. Traditional spatial interpolation methods, such as Kriging and IDW, are affected by sensor deployment density and positioning accuracy, and are prone to oversmoothing problems in sparse monitoring networks. Typical machine learning techniques, such as GANs and GNNs, show strong sensitivity to the spatial distribution of sampling points and the presence of measurement noise. Most existing research focuses on two-dimensional REM construction. With the development of integrated air-air-ground networks, the spectral situation has expanded from two-dimensional space to three-dimensional space. A single horizontal slice of a two-dimensional spectrum map can no longer accurately describe the electromagnetic situation in three-dimensional space. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing technologies, this invention provides a three-dimensional spectrum situation completion method for complex urban electromagnetic environments. It can complete the three-dimensional spectrum situation based on sparse monitoring data and urban map information, providing a data foundation for cognitive radio spectrum sharing, base station layout and site selection, and the construction of three-dimensional radio environment maps in low-altitude scenarios.

[0005] Technical Solution: A three-dimensional spectral situational awareness method for complex urban electromagnetic environments, comprising the following steps:

[0006] Acquire complete 3D spectral situational training data organized by horizontal grid and height layers;

[0007] A three-dimensional building mask and a validity mask for identifying the sampleable grid are generated based on the spatial distribution of buildings in the target area. A two-dimensional sampling mask is generated in the sampleable grid on the horizontal plane. The two-dimensional sampling mask is broadcast along the height dimension and combined with the validity mask to obtain the three-dimensional sampling mask and its corresponding sparse three-dimensional spectral situation.

[0008] Feature extraction and intra-layer fusion are performed on sparse 3D spectral situation, 3D sampling mask and 3D building mask at the same height layer, respectively. In addition, cross-height layer fusion is performed on the intra-layer fused features of different height layers to obtain conditional features.

[0009] Noise is added to the complete three-dimensional spectrum situation training data, and the added noise is used as the supervision target. A noise prediction network is trained based on conditional features and the complete three-dimensional spectrum situation after noise addition.

[0010] The actual sparse 3D spectral situation, actual sampling location and building spatial distribution of the target area are processed in the same conditional feature generation method as in the training phase. Starting with the random noise tensor, the trained noise prediction network is used to perform inverse denoising under the conditional feature constraints to output the complete 3D spectral situation.

[0011] Furthermore, complete 3D spectral situational training data organized by horizontal grid and height layers is obtained, including:

[0012] Obtain historical complete three-dimensional spectral situation data of the target area, or use simulation software to obtain empirical complete three-dimensional spectral situation data of the target area;

[0013] The complete three-dimensional spectral situation data is divided into horizontal grids according to the preset horizontal grid resolution, and multiple height layers are selected in the vertical direction to form a complete three-dimensional spectral situation tensor organized with two horizontal dimensions and one height dimension.

[0014] Multiple complete three-dimensional spectral situation tensors are combined to obtain complete three-dimensional spectral situation training data.

[0015] Furthermore, a 3D building mask and a validity mask are generated based on the spatial distribution of buildings in the target area, including:

[0016] Extract the spatial distribution of buildings from the 3D map information of the target area, and perform gridding on the spatial distribution of buildings according to the horizontal grid and height layer corresponding to the complete 3D spectrum situation training data.

[0017] The grids occupied by buildings and those not occupied by buildings in each height layer are distinguished and marked to obtain the 3D building mask;

[0018] The validity of each mesh is determined based on the 3D building mask. Mesh not occupied by buildings is marked as valid mesh, and mesh occupied by buildings is marked as invalid mesh, thus obtaining the validity mask.

[0019] Furthermore, the three-dimensional sampling mask and its corresponding sparse three-dimensional spectral state are obtained, including:

[0020] Based on a preset sampling rate, a horizontal sampling position is randomly selected on the horizontal plane, and a two-dimensional sampling mask is generated based on the selected horizontal sampling position.

[0021] The two-dimensional sampling mask is broadcast along the height dimension to each height layer to obtain the broadcast sampling mask. The broadcast sampling mask is multiplied element-wise with the validity mask to exclude the grid occupied by buildings in each height layer to obtain the three-dimensional sampling mask.

[0022] The complete three-dimensional spectral situation training data is masked by a three-dimensional sampling mask to obtain a sparse three-dimensional spectral situation.

[0023] Furthermore, the conditional features are obtained, including:

[0024] For each height layer, the sparse 3D spectral situation, 3D sampling mask and 3D building mask of that height layer are stacked into a three-channel tensor, and then feature extraction is performed through independent convolutional blocks to obtain the intra-layer fusion features of the corresponding height layer.

[0025] The intra-layer fusion features of different height layers are spliced ​​along the channel dimension, and the spliced ​​features are sequentially subjected to channel compression and spatial feature extraction to obtain conditional features that fuse information from different height layers and match the number of noise feature channels with the complete three-dimensional spectral situation after noise addition.

[0026] Furthermore, the noise prediction network comprises a structured, highly fused subnetwork and a U-Net backbone network;

[0027] The structured, highly fused subnetwork is used to perform intra-layer fusion and cross-height layer fusion to obtain conditional features;

[0028] Convolutional feature extraction is performed on the complete three-dimensional spectrum situation after noise addition to obtain noise features. The conditional features and noise features are added element by element to obtain fused features, and the fused features are input into the U-Net backbone network.

[0029] The U-Net backbone network consists of an encoder, a bottleneck layer, and a decoder. The encoder and decoder each include multiple resolution levels, and an attention mechanism is set at at least one resolution level.

[0030] Furthermore, the noise prediction network is trained, including:

[0031] A complete three-dimensional spectrum situation is selected from the complete three-dimensional spectrum situation training data, and Gaussian noise is gradually added to the selected complete three-dimensional spectrum situation to form a noisy three-dimensional spectrum situation in the forward diffusion process.

[0032] The conditional features corresponding to the noisy 3D spectral situation, the noisy 3D spectral situation, and the diffusion time step are input into the noise prediction network to obtain the predicted noise.

[0033] The training loss is determined based on the mean square error between the predicted noise and the Gaussian noise added to the complete three-dimensional spectral situation. The network parameters of the noise prediction network are then iteratively adjusted based on the training loss until the training loss meets the convergence condition.

[0034] Furthermore, the trained noise prediction network is used to perform inverse denoising under conditional feature constraints, including:

[0035] Generate a Gaussian random noise tensor with the same spatial shape as the complete three-dimensional spectral situation to be output, and use the Gaussian random noise tensor as the initial noisy tensor for inverse denoising;

[0036] The diffusion time steps are decreased according to the preset step size. At each diffusion time step, the current noisy tensor, the current diffusion time step, and the conditional features are input into the trained noise prediction network to obtain the predicted noise at the current diffusion time step.

[0037] Based on the current noisy tensor, the predicted noise, and the cumulative noise retention ratio corresponding to the current diffusion time step, determine the estimated value of the denoised data for the current diffusion time step.

[0038] Before the termination time step is reached, the noisy tensor for the next diffusion time step is generated based on the estimated value of the denoised data, the predicted noise, and the cumulative noise retention ratio corresponding to the next diffusion time step. When the termination time step is reached, the obtained denoised data is output as the complete three-dimensional spectral situation.

[0039] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the above-described method for three-dimensional spectral situational awareness in complex urban electromagnetic environments.

[0040] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for three-dimensional spectral situation completion for complex urban electromagnetic environments.

[0041] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for three-dimensional spectral situational awareness in complex urban electromagnetic environments.

[0042] Beneficial effects: This invention determines the sampleable grid based on the spatial distribution of buildings in the target area, generates a two-dimensional sampling mask on the horizontal plane, and broadcasts it along the height dimension to form a three-dimensional sampling constraint, so that the sparse observation location matches the building distribution and actual acquisition method in the complex urban environment; by performing intra-layer fusion of sparse spectral status, sampling mask and building mask at each height layer, and performing cross-height layer fusion of intra-layer fusion features at different height layers, it is possible to extract the horizontal spatial features of each height plane and the vertical correlation features between different height layers respectively; furthermore, the formed conditional features are fused with noise features, so that the actual sparse observation information and building spatial constraints directly participate in the noise prediction and reverse denoising process, thereby achieving three-dimensional completion of missing spectral status data in complex urban electromagnetic environments without relying on transmitter position and power information. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 The RMSE index of the example is compared with that of other typical methods;

[0045] Figure 3 This is a visual comparison of the embodiments with other typical methods;

[0046] Figure 4 The following is a comparison chart of the completion effects for different complexity scenarios in the example. Detailed Implementation

[0047] To provide a clearer understanding of the features and advantages of the technical solution of the present invention, the composition and implementation of the specific solution are described below in conjunction with the accompanying drawings.

[0048] Reference Figure 1 This invention proposes a three-dimensional spectral situation completion method for complex urban electromagnetic environments, specifically including the following steps:

[0049] Step 1: Obtain several historical or empirical radio acquisition data through radio detection equipment and preprocess them to construct a complete three-dimensional spectrum situation training dataset of a certain scale. The complete three-dimensional spectrum situation training dataset contains multiple height layers.

[0050] Considering cost, historical complete 3D spectrum situational data of the target area is obtained. This data comprises a 3D tensor composed of radio propagation parameters such as the received signal strength (RSS) at each spatial grid location and at each altitude level within the target area. Alternatively, empirical complete 3D spectrum situational data of the target area can be obtained using simulation software WinProp. WinProp can calculate the received signal strength at each spatial location within the target area based on transmitter parameters, building models, and propagation models, thereby generating complete 3D spectrum situational data. The complete 3D spectrum situational data is then gridded, with a grid resolution in the horizontal direction. The process involves selecting representative height layers in the vertical direction to obtain a complete three-dimensional spectral situation tensor, denoted as... Where W and L are the number of grids divided in the horizontal longitude and latitude directions, respectively, and N is the number of height layer slices divided in the vertical direction.

[0051] Following these steps, several complete three-dimensional spectral situation data are processed to obtain a certain amount of complete three-dimensional spectral situation tensors.

[0052] Step 2: Obtain the spatial distribution information of buildings based on the map information of the 3D target area, and generate building masks layer by layer at each height to represent complex 3D environment information.

[0053] By using publicly available 3D target area map information, building occlusion information of the 3D target area is extracted, and then processed according to the horizontal plane grid resolution from the first step. The number of slices at different height levels is used to perform gridding on the acquired 3D map information, which includes 2D outline information of buildings and building height information.

[0054] Obtain the 3D map's planar and height information, mark the horizontal grid positions of buildings at each height level as inaccessible areas, assigning a value of 1 to their corresponding grid points, and assigning a value of 0 to grid points in areas not containing buildings. This yields the 3D building mask. .

[0055] Step 3: Perform random sparse sampling on the area to be tested through hierarchical random sampling to obtain sparse sampled spectrum situation data and sampling mask.

[0056] Considering that the deployment location of the measuring equipment will be affected by factors such as building distribution and accessibility, based on the three-dimensional building mask B obtained in the second step, the non-building areas in the target area are sampled according to the sampling rate r, and the number of points to be sampled is counted.

[0057] At the same horizontal coordinate, the received signal strength at different heights is simultaneously collected by the same measuring sensor. Therefore, a two-dimensional mask broadcasting mechanism is adopted, first on the horizontal plane... Random sampling is performed to generate a sampling mask, which is then extended along the height dimension to all N layers. Specifically, the two-dimensional sampling mask randomly selects measurement points on the horizontal plane based on the sampleable locations. The sampling mask value of the selected grid positions is set to 1. This process is then extended along the height dimension to all N layers, with non-building areas designated as valid masks. The sampling mask and validity mask after broadcasting Element-wise multiplication is performed to obtain sparse sampled spectral situational data, satisfying the required number of sampling points. and 3D sampling mask .

[0058] This invention proposes a three-dimensional hierarchical sampling mechanism for practical spectrum measurement processes. The effective sampling area is determined based on the spatial distribution of buildings, a sampling mask is generated on the horizontal plane, and the mask is extended along the height direction to form a three-dimensional sampling constraint. This mechanism ensures that the sampling process is consistent with the three-dimensional spatial structure of the city and guarantees the spatial correlation between sampling positions at different height layers, providing conditions and constraints that conform to the actual measurement laws for subsequent diffusion completion.

[0059] Step 4: Use the sampling situation data and sampling mask obtained in Step 3, as well as the building mask obtained in Step 2, as the conditional input.

[0060] The building mask obtained in the second step The sampled spectrum situational data obtained in the third step and the corresponding three-dimensional sampling mask All three have the same shape. Use these three items together as conditional input. .

[0061] Step 5: Input the conditional input obtained in Step 4 into the denoising diffusion model for training. Set the learning objective as the complete three-dimensional spectral situation in Step 1, and train to obtain the U-Net noise prediction network that has been trained.

[0062] 5.1) Construct a training set, combining the complete three-dimensional spectral situational data from step one with the three modalities of the conditional input after transformation of each group of samples from step four. , and Combine samples to construct a sample, and create several samples to obtain a training set.

[0063] 5.2) Define the diffusion process by progressively adding Gaussian noise to the complete spectral situational data from the first step, forming a forward Markov noise sequence. Specifically, at each time... At this point, the process is defined as:

[0064] ,

[0065] in, The total number of diffusion steps, The noise schedule is a gradually decreasing signal-to-noise ratio. Let it be the identity matrix. Usually, let... For a random A monotonically increasing small sequence causes noise to accumulate and increase with each step, until... hour, It approximately follows a standard Gaussian distribution. .

[0066] 5.3) Construct a denoising prediction network, which consists of a structured high-fusion submodule and a U-Net backbone submodule: The structured high-fusion submodule first performs intra-layer fusion. For each high-level layer, it selects three modalities: the sampled spectral situation data, the sampled mask, and the building mask, and stacks them in the channel dimension to form a three-channel tensor. Then through independent convolutional blocks Extracting local spatial features of each layer The process involves obtaining intra-layer features of a preset dimension; then, the intra-layer features extracted from each height layer are concatenated along the channel direction, and feature compression and fusion are performed through convolution operations to achieve cross-height layer information interaction, i.e., inter-layer fusion, and finally outputting conditional features that match the number of noise feature channels. Simultaneously, convolutional feature extraction is performed on the noisy spectral situation data obtained in step 5.2 to obtain noise features. The two are added together to obtain , As input to the U-Net backbone network, the first convolutional layer of the U-Net backbone network outputs features with a preset number of channels. The encoder gradually increases the number of feature channels and introduces an attention module at a preset level. The encoder and decoder each have several residual blocks, and the bottleneck layer has residual blocks to further extract deeper features. Parameters such as feature dimension, convolutional kernel size, network depth, and number of channels can be adjusted according to specific task requirements. In this embodiment, the first convolutional layer of the U-Net backbone network outputs 128 channels. The number of channels in each layer of the encoder is successively increased to [1, 2, 2, 2] times the number of input channels. An attention mechanism is introduced in the 2nd and 3rd layers. The encoder has 2 residual blocks at each resolution level, the decoder has 3 residual blocks at each resolution level, and the bottleneck layer consists of 2 residual blocks. After training, this network can be used for noise prediction.

[0067] 5.4) The fusion features obtained in step 5.3 Input to U-Net backbone network, output predicted noise and with real noise Calculate the mean square error loss ;

[0068] 5.5) Repeat steps 5.1) to 5.4) iteratively optimize the network parameters until the loss function converges, finally obtaining the trained U-Net noise prediction network. .

[0069] Unlike existing technologies that use 3D convolution to extract features from spatial volume data uniformly, this invention takes into account the differences in the impact of building obstruction, propagation path and spatial accessibility on different height layers during urban electromagnetic propagation. It explicitly models the height dimension as a physically meaningful structural dimension, enabling the model to extract the horizontal propagation features of each height plane and further learn the vertical propagation correlation between different heights, thereby improving the ability to complete the spectrum situation in complex 3D urban environments.

[0070] Step 6: Perform random and sparse monitoring sampling on the current three-dimensional target area, collect actual radio monitoring data and preprocess it to obtain the actual data of the three-dimensional target area's complete spectrum situation.

[0071] Radio detection equipment such as drones and sensors are used to collect spectrum data of a three-dimensional target area. Random sampling is performed according to the sampling rate to obtain actual three-dimensional spectrum situation sampling data. After processing abnormal data, the sampled spectrum situation is preprocessed according to the preprocessing method in the first step.

[0072] Step 7: Use the sparse spectrum situation to be completed, the sampling mask, and the building mask obtained in Step 6 as conditional inputs. The three modes, after being highly structured and fused, yield conditional features characterizing the actual sampled spectrum situation. .

[0073] Step 8: Apply the conditional features obtained in Step 7 The features extracted by convolution with random noise are added to obtain fused features, which are then fed into the trained denoising diffusion probability model to generate fully restored three-dimensional spectral situation data and draw a complete three-dimensional radio map.

[0074] 8.1) Generate a Gaussian random noise tensor with the same three-dimensional spatial shape as the spatial shape of the complete situation map corresponding to the input conditions. This serves as the starting point for the reverse denoising process, and the diffusion process is set to require a total of iterations. step;

[0075] 8.2) Let the time step To begin, As a termination condition, proceed with a fixed step size. By sequentially decreasing the noise level, each step executes the following three sub-steps to gradually reconstruct a low-noise version of the spectral situation data:

[0076] 8.2.1) Convert the noisy tensor corresponding to the current round. The conditions obtained at iteration time step t and in step 7 are fed together into the already trained U-Net noise prediction network. In the process, the network outputs the noise estimate required for this round. ;

[0077] 8.2.2) The noise estimate obtained in the previous step With the current noise tensor Substitute into the formula This will give you the corresponding denoised data estimate. ,in This represents the cumulative result after multiplying the noise retention ratios of each round step by step.

[0078] 8.2.3) Once the iteration time step of the current round... Based on the advancement formula of the denoising diffusion implicit model DDIM algorithm The noisy result corresponding to the next round is calculated. ;

[0079] 8.3) When the iteration proceeds to After the final update is completed, the denoised data obtained at this time is... The final result is a complete three-dimensional spectral situation map.

[0080] This invention addresses the spectral situation recovery problem under sparse observation constraints. It does not rely on any transmitter location or power information; its conditional information primarily comes from sparse RSS observations and publicly available map information. The model injects known observation information into the diffusion denoising process through a conditional input mechanism, making the generation process constrained by the real observation area. During the training phase, it learns the mapping relationship between conditional observations and the complete spectral situation. During the inference phase, it recovers the missing region solely through a diffusion backsampling process, without requiring additional gradient optimization steps. This offers advantages in computational efficiency and constraint stability.

[0081] To verify the effectiveness of this invention, the Altair WinProp electromagnetic simulation platform was used to generate 3D simulated spectral situational data based on the Dominant Path Model (DPM), constructing urban building background scenes of low, medium, and high complexity. The proposed method is named REM-Diff-3D. Baseline methods for comparison include the traditional spatial interpolation method Kriging, and two deep learning methods, Autoencoder and cGAN. Simulation results are shown below. Figure 2 , Figure 3 and Figure 4All methods were trained and tested on the same dataset, with the same random sparse sampling mask and building mask conditions, and the root mean square error (RMSE) was used as the core evaluation metric.

[0082] like Figure 2 As shown, within the sampling rate range of 5% to 30%, the REM-Diff-3D method proposed in this invention consistently maintains the lowest reconstruction error. It is worth noting that the reconstruction accuracy of REM-Diff-3D under extremely low sampling rate conditions is superior to that of other baseline methods at the highest sampling rate, which fully demonstrates the robustness advantage of the structured high fusion and conditional diffusion mechanism proposed in this invention under extremely sparse observation conditions.

[0083] like Figure 3 As shown, (a), (b), (c), and (d) represent the visual effects of REM-Diff-3D, cGAN, Kriging, and Autoencoder, respectively. Qualitative comparison results show that Kriging struggles to characterize signal abrupt changes in building-occluded areas. Autoencoder, due to the compression loss of spatial details caused by bottleneck structures, struggles to distinguish between shadowed and open areas. While cGAN can recover some spatial details through adversarial training, it is prone to noise and structural artifacts on cross-sections at various height levels. In contrast, REM-Diff-3D exhibits good reconstruction fidelity at all height levels, and the signal intensity variation trend between different height levels matches the actual physical laws of electromagnetic propagation. This indicates that the proposed method not only learns spatial correlations within the horizontal plane but also effectively models the signal propagation characteristics in the vertical direction.

[0084] like Figure 4 As shown, (a), (b), and (c) represent the 3D spectral situation completion effects in urban backgrounds with low, medium, and high complexity, respectively. It can be seen that under the same sampling rate conditions in urban scenes with low, medium, and high complexity, the performance difference of REM-Diff-3D is small among different complexity scenes, indicating that the method proposed in this invention has good adaptability and cross-scene generalization ability to changes in urban environmental complexity.

[0085] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the three-dimensional spectral situation completion method for complex urban electromagnetic environments as described above.

[0086] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the three-dimensional spectral situation completion method for complex urban electromagnetic environments as described above.

[0087] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the three-dimensional spectral situation completion method for complex urban electromagnetic environments as described above.

Claims

1. A three-dimensional spectral situational awareness method for complex urban electromagnetic environments, characterized in that, Includes the following steps: Acquire complete 3D spectral situational training data organized by horizontal grid and height layers; A three-dimensional building mask and a validity mask for identifying the sampleable grid are generated based on the spatial distribution of buildings in the target area. A two-dimensional sampling mask is generated in the sampleable grid on the horizontal plane. The two-dimensional sampling mask is broadcast along the height dimension and combined with the validity mask to obtain the three-dimensional sampling mask and its corresponding sparse three-dimensional spectral situation. Feature extraction and intra-layer fusion are performed on sparse 3D spectral situation, 3D sampling mask and 3D building mask at the same height layer, respectively. In addition, cross-height layer fusion is performed on the intra-layer fused features of different height layers to obtain conditional features. Noise is added to the complete three-dimensional spectrum situation training data, and the added noise is used as the supervision target. A noise prediction network is trained based on conditional features and the complete three-dimensional spectrum situation after noise addition. The actual sparse 3D spectral situation, actual sampling location and building spatial distribution of the target area are processed in the same conditional feature generation method as in the training phase. Starting with the random noise tensor, the trained noise prediction network is used to perform inverse denoising under the conditional feature constraints to output the complete 3D spectral situation.

2. The method according to claim 1, characterized in that, Acquire complete 3D spectral situational training data organized by horizontal grid and height layers, including: Obtain historical complete three-dimensional spectral situation data of the target area, or use simulation software to obtain empirical complete three-dimensional spectral situation data of the target area; The complete three-dimensional spectral situation data is divided into horizontal grids according to the preset horizontal grid resolution, and multiple height layers are selected in the vertical direction to form a complete three-dimensional spectral situation tensor organized with two horizontal dimensions and one height dimension. Multiple complete three-dimensional spectral situation tensors are combined to obtain complete three-dimensional spectral situation training data.

3. The method according to claim 1, characterized in that, Generate a 3D building mask and a validity mask based on the spatial distribution of buildings in the target area, including: Extract the spatial distribution of buildings from the 3D map information of the target area, and perform gridding on the spatial distribution of buildings according to the horizontal grid and height layer corresponding to the complete 3D spectrum situation training data. The grids occupied by buildings and those not occupied by buildings in each height layer are distinguished and marked to obtain the 3D building mask; The validity of each mesh is determined based on the 3D building mask. Mesh not occupied by buildings is marked as valid mesh, and mesh occupied by buildings is marked as invalid mesh, thus obtaining the validity mask.

4. The method according to claim 3, characterized in that, The three-dimensional sampling mask and its corresponding sparse three-dimensional spectral state are obtained, including: Based on a preset sampling rate, a horizontal sampling position is randomly selected on the horizontal plane, and a two-dimensional sampling mask is generated based on the selected horizontal sampling position. The two-dimensional sampling mask is broadcast along the height dimension to each height layer to obtain the broadcast sampling mask. The broadcast sampling mask is multiplied element-wise with the validity mask to exclude the grid occupied by buildings in each height layer to obtain the three-dimensional sampling mask. The complete three-dimensional spectral situation training data is masked by a three-dimensional sampling mask to obtain a sparse three-dimensional spectral situation.

5. The method according to claim 1, characterized in that, The conditional features obtained include: For each height layer, the sparse 3D spectral situation, 3D sampling mask and 3D building mask of that height layer are stacked into a three-channel tensor, and then feature extraction is performed through independent convolutional blocks to obtain the intra-layer fusion features of the corresponding height layer. The intra-layer fusion features of different height layers are spliced ​​along the channel dimension, and the spliced ​​features are sequentially subjected to channel compression and spatial feature extraction to obtain conditional features that fuse information from different height layers and match the number of noise feature channels with the complete three-dimensional spectral situation after noise addition.

6. The method according to claim 1, characterized in that, The noise prediction network consists of a structured, highly integrated subnetwork and a U-Net backbone network; The structured, highly fused subnetwork is used to perform intra-layer fusion and cross-height layer fusion to obtain conditional features; Convolutional feature extraction is performed on the complete three-dimensional spectrum situation after noise addition to obtain noise features. The conditional features and noise features are added element by element to obtain fused features, and the fused features are input into the U-Net backbone network. The U-Net backbone network consists of an encoder, a bottleneck layer, and a decoder. The encoder and decoder each include multiple resolution levels, and an attention mechanism is set at at least one resolution level.

7. The method according to claim 1, characterized in that, Training the noise prediction network includes: A complete three-dimensional spectrum situation is selected from the complete three-dimensional spectrum situation training data, and Gaussian noise is gradually added to the selected complete three-dimensional spectrum situation to form a noisy three-dimensional spectrum situation in the forward diffusion process. The conditional features corresponding to the noisy 3D spectral situation, the noisy 3D spectral situation, and the diffusion time step are input into the noise prediction network to obtain the predicted noise. The training loss is determined based on the mean square error between the predicted noise and the Gaussian noise added to the complete three-dimensional spectral situation. The network parameters of the noise prediction network are then iteratively adjusted based on the training loss until the training loss meets the convergence condition.

8. The method according to claim 1, characterized in that, The trained noise prediction network is used to perform inverse denoising under conditional feature constraints, including: Generate a Gaussian random noise tensor with the same spatial shape as the complete three-dimensional spectral situation to be output, and use the Gaussian random noise tensor as the initial noisy tensor for inverse denoising; The diffusion time steps are decreased according to the preset step size. At each diffusion time step, the current noisy tensor, the current diffusion time step, and the conditional features are input into the trained noise prediction network to obtain the predicted noise at the current diffusion time step. Based on the current noisy tensor, the predicted noise, and the cumulative noise retention ratio corresponding to the current diffusion time step, determine the estimated value of the denoised data for the current diffusion time step. Before the termination time step is reached, the noisy tensor for the next diffusion time step is generated based on the estimated value of the denoised data, the predicted noise, and the cumulative noise retention ratio corresponding to the next diffusion time step. When the termination time step is reached, the obtained denoised data is output as the complete three-dimensional spectral situation.

9. An electronic device, characterized in that, It includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the three-dimensional spectral situation completion method for complex urban electromagnetic environments as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the three-dimensional spectrum situation completion method for complex urban electromagnetic environments as described in any one of claims 1-8.

Images