Method for constructing three-dimensional spectrum situation based on autonomous planning of unmanned aerial vehicle

Abstract

The application discloses a three-dimensional spectrum situation construction method based on unmanned aerial vehicle autonomous planning, comprising the following steps: step 1, three-dimensional modeling and grid division are performed on a low-altitude target area to obtain a three-dimensional discrete space representation of the target area; step 2, received signal strength is collected to form sparse spectrum observation data; step 3, input data of a spectrum situation large model is constructed; step 4, the input data is subjected to feature extraction and three-dimensional spectrum situation reconstruction by using a spectrum situation large model based on a Transform structure to obtain an initial estimation result of the three-dimensional spectrum situation; step 5, the initial estimation result is subjected to physical perception correction; step 6, a next sampling position is autonomously determined and a sampling track is updated; step 7, the unmanned aerial vehicle continues to collect new spectrum data and performs iterative reconstruction to output a complete three-dimensional spectrum situation map. The application can improve the three-dimensional spectrum situation construction precision and the unmanned aerial vehicle sampling efficiency under the condition of a low sampling ratio and is suitable for low-altitude network spectrum sensing, interference analysis and wireless network planning and the like.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a method for constructing a three-dimensional spectrum situation based on autonomous planning by unmanned aerial vehicles (UAVs). Background Technology

[0002] With the rapid development of 6G mobile communication, low-altitude intelligent networks, and integrated air-space-ground networks, the number of drones, low-altitude base stations, sensing terminals, and various wireless devices in low-altitude airspace is constantly increasing, significantly enhancing the utilization density of electromagnetic spectrum resources. Compared with traditional terrestrial wireless networks, the electromagnetic propagation environment in low-altitude networks is more complex. On the one hand, mobile platforms such as drones are characterized by rapid altitude changes, wide spatial distribution, and diverse service models; on the other hand, factors such as urban buildings, terrain undulations, and irregular obstacles can cause obstruction, shadow fading, and multipath propagation, resulting in a significant three-dimensional non-uniformity in signal strength distribution within the target area. Spectrum situational awareness construction, as a crucial foundation for spectrum sensing, interference analysis, resource allocation, and low-altitude communication network planning, requires accurate, continuous, and visual characterization of the electromagnetic field strength distribution within the target area. Therefore, how to efficiently construct a low-altitude three-dimensional spectrum situational map under limited sampling conditions has become a key issue in intelligent spectrum management.

[0003] Existing methods for constructing spectrum situations mainly include traditional interpolation methods, deep learning methods, and UAV-assisted sampling methods. Traditional methods typically utilize Kriging interpolation, inverse distance weighted interpolation, K-nearest neighbors, or tensor completion to recover the spectrum distribution of unknown regions from a small number of observation points. In recent years, deep learning methods have been gradually introduced into radio mapping and spectrum situation construction tasks. For example, Y. Teganya and D. Romero et al. proposed a radio map estimation method based on deep completion autoencoders in their paper "Deep completion autoencoders for radio map estimation" (IEEE Transactions on Wireless Communications, vol. 21, no. 3, pp. 1710–1724, Mar. 2022), which recovers a complete radio map from sparse observation data through a neural network. Compared with traditional interpolation methods, this type of method has stronger nonlinear modeling capabilities and can learn complex spatial dependencies to a certain extent. However, its performance still depends on sampling density and training data quality. When the observation points are extremely sparse or the sampling locations are irregular, the model reconstruction accuracy is prone to decline, and its adaptability to physical factors such as building occlusion, multipath propagation and height changes in low-altitude scenes is still limited.

[0004] For the problem of low-altitude 3D spectrum sampling, unmanned aerial vehicles (UAVs) are considered an important carrier for constructing low-altitude spectrum situation maps due to their flexible deployment, controllable flight altitude, and maneuverable sampling capabilities. For example, Q. Wu, F. Shen, Z. Wang et al. proposed a UAV 3D spectrum mapping method based on region of interest (ROI-driven UAV deployment) in their paper "3D spectrum mapping based on ROI-driven UAV deployment" (IEEE Network, vol.34, no. 5, pp. 24–31, Sep. / Oct. 2020), which improves the sampling efficiency of key areas by optimizing the deployment location of UAVs. Furthermore, Y. Chen, Q. Zhu, J. Wang et al. studied an UAV-aided efficient informative path planning for autonomous 3D spectrum mapping in their paper "UAV-aided efficient informative path planning for autonomous 3D spectrum mapping" (IEEE Transactions on Cognitive Communications and Networking, 2026, early access.), which improves the efficiency of 3D spectrum sampling through target point selection and local path optimization. However, most of the above methods rely on preset trajectories, fixed sampling rules, or local path optimization strategies, and their ability to proactively discover high-uncertainty areas in unknown environments remains insufficient.

[0005] With the development of large-scale model and agent technologies, learning electromagnetic propagation laws using large-scale models and combining them with autonomous agent interaction for sampling decision-making provides a new technical path for low-altitude spectrum situational awareness construction. For example, H. Quan, W. Ni, T. Zhang et al. proposed a radio map generation and wireless network planning method based on large language model agents in their paper "Large language model agents for radio map generation and wireless network planning" (IEEE Network Letters, vol. 7, no. 3, pp. 166–170, Sept. 2025), verifying the application potential of large-scale model agents in automated wireless network planning. Nevertheless, existing large-scale model-related methods focus more on wireless network planning or two-dimensional radio map generation, and still lack integrated solutions for problems such as sparse sampling, building occlusion correction, and UAV autonomous trajectory updates in low-altitude three-dimensional scenes.

[0006] In summary, existing technologies still have the following shortcomings: First, traditional interpolation and tensor completion methods make strong assumptions about local smoothness, making it difficult to accurately recover signal abrupt change regions in complex low-altitude obstruction environments. Second, while conventional deep learning methods have strong data fitting capabilities, they lack stability under extremely low sampling ratios and irregular sampling conditions, and do not fully utilize electromagnetic propagation mechanisms. Third, existing UAV spectrum sampling methods often lack closed-loop linkage with the spectrum reconstruction model, making it difficult to proactively select high-value sampling locations based on the uncertainty of the current reconstruction results. Therefore, it is urgent to propose a UAV autonomous sampling and large-model 3D spectrum situation construction method for low-altitude networks, organically combining 3D spectrum reconstruction, uncertainty-aware sampling, and physical propagation correction to achieve high-precision and robust low-altitude 3D spectrum situation construction under limited sampling budgets. Summary of the Invention

[0007] The technical problem this invention addresses is the insufficient reconstruction accuracy, low UAV sampling efficiency, and weak physical consistency of existing low-altitude spectrum situational awareness construction methods under sparse sampling conditions. This invention proposes a UAV autonomous sampling and large-model 3D spectrum situational awareness construction method for low-altitude networks. This method constructs a low-altitude 3D gridded spectrum representation and introduces a large-scale spectrum situational awareness model with multi-scale spatial feature extraction capabilities to recover the complete 3D spectrum distribution of the target region from sparse observation data. Simultaneously, by combining environmental information such as building distribution, height information, and sampling masks, the model can better characterize occlusion, shadow fading, and multipath propagation characteristics in low-altitude scenarios. Furthermore, this invention utilizes an uncertainty perception mechanism to evaluate regions with insufficient information in the current spectrum reconstruction results and autonomously plans the next sampling position by comprehensively considering the degree of spectrum spatial variation and the UAV flight cost, thus forming a closed-loop process of "sampling—reconstruction—correction—resampling." In addition, this invention compensates for the reconstruction results in non-line-of-sight occlusion areas through a physical perception correction mechanism, improving the accuracy and stability of 3D spectrum situational awareness construction in complex building environments, providing an efficient and reliable technical solution for low-altitude network spectrum perception, interference analysis, and intelligent spectrum management.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for constructing a three-dimensional spectrum situation based on UAV autonomous planning includes:

[0010] Step 1: Construct a three-dimensional spatial model of the low-altitude target area and divide the target area into grids to obtain a three-dimensional discrete spatial representation for spectrum situation construction;

[0011] Step 2: The UAV performs spectrum sampling according to the initial trajectory within the low-altitude target area to obtain the received signal strength at different spatial locations, forming sparse spectrum observation data;

[0012] Step 3: Based on the sparse spectrum observation data, sampling locations, building distribution information, and height information, construct the input data for the large spectrum situation model;

[0013] Step 4: Using a large-scale spectral situation model based on the Transformer structure, feature extraction and three-dimensional spectral situation reconstruction are performed on the input data to obtain the initial estimation results of the three-dimensional spectral situation of the target area;

[0014] Step 5: Based on the building occlusion relationship and the observed spectrum data, perform physical perception correction on the initial estimation result of the three-dimensional spectrum situation to obtain a three-dimensional spectrum situation map with enhanced physical consistency;

[0015] Step 6: Based on the uncertainty of the three-dimensional spectrum situation map, combined with the degree of spectrum spatial change and the flight cost of the UAV, determine the next sampling position and update the UAV sampling trajectory;

[0016] Step 7: The UAV continues to collect spectrum data according to the updated sampling trajectory, and inputs the newly sampled data into the spectrum situation large model for iterative reconstruction until the preset sampling budget or reconstruction accuracy requirements are met, and outputs a complete three-dimensional spectrum situation map.

[0017] To optimize the above technical solution, the specific measures also include:

[0018] Step 1 above specifically includes:

[0019] Representing the low-altitude target region as a three-dimensional spatial region: And divide it into: A three-dimensional mesh, in which , , These represent the number of grid cells in the horizontal, vertical, and vertical directions of the target region, respectively; let the actual size of the target region be... , , Then the grid spacing in each direction is:

[0020] ,

[0021] in, , , .

[0022] Step 2 above specifically includes:

[0023] Step 201: The drone's spatial location and frequency The received signal strength acquired below is expressed as:

[0024] ,

[0025] in, Indicates the received signal strength. Indicates the number of launch sources. Indicates the first Each transmitter is at frequency The transmit power at the following levels Indicates the corresponding channel gain. This indicates noise and interference.

[0026] Step 202: The UAV collects the received signal strength at several locations along its flight path to form a sparse spectrum observation dataset. ,in, This represents a sparse spectral observation dataset. This indicates the number of sampled points.

[0027] Step 3 above specifically includes:

[0028] Step 301: Construct a multimodal input tensor by using the sampled spectrogram, sampling mask, building distribution information, and height information as model inputs: in, Indicates the number of input channels.

[0029] Step 302: To facilitate model processing, the height dimension and channel dimension are merged to obtain the transformed input tensor, resulting in: This processing method can reduce the computational complexity of the model on 3D data while preserving height information.

[0030] Step 4 above specifically includes:

[0031] Step 401: Use a large spectrum situation model based on the Transformer structure as the reconstruction network. The large spectrum situation model includes a Patch embedding module, a hierarchical encoder, a trajectory-aware encoder, a hierarchical decoder, and a jump connection module.

[0032] Step 402: The input tensor is divided into several local blocks by the Patch embedding module and mapped to a high-dimensional feature space; spatial correlation features at different scales are extracted by the hierarchical encoder; the UAV sampling trajectory and sampling mask information are fused by the trajectory-aware encoding module; the spatial resolution is restored by the hierarchical decoder, and the three-dimensional spectral situation estimation result is output.

[0033] Step 403: The three-dimensional spectral situation reconstruction process is represented as follows: .in, This represents the reconstructed three-dimensional spectral situation map. Indicates by parameters A large model representing the spectral situation.

[0034] Step 404: During model training, the reconstruction error is calculated for the unobserved region, and the loss function is expressed as:

[0035] ,

[0036] in, Represents the set of unobserved locations. Indicates position The actual spectral value at that location, Indicates position The reconstructed spectral value at that location.

[0037] Step 5 above specifically includes:

[0038] Step 501: Based on the building distribution information and the estimated location of the emission source, determine whether there is occlusion between the emission source and each grid location, and construct a line-of-sight propagation discrimination map:

[0039] ,

[0040] in, This indicates that the location is within the line-of-sight propagation area. This indicates that the location is a non-line-of-sight occlusion area. For non-line-of-sight areas, the prediction bias is estimated based on the observed samples:

[0041] ,

[0042] in, This indicates the number of samples observed in non-line-of-sight regions. This represents the average prediction bias in non-line-of-sight areas.

[0043] Step 502: Correct the non-line-of-sight region using the average prediction bias to obtain a three-dimensional spectral situation map with enhanced physical consistency:

[0044] ,

[0045] in, This represents the spectral state value after physical perception correction.

[0046] Step 6 above specifically includes:

[0047] Step 601: Perform multiple random forward propagations during the model inference phase to obtain multiple three-dimensional spectral situation prediction results, and calculate the uncertainty graph based on the variance between different prediction results. .

[0048] Step 602: When determining the next sampling location, a sampling scoring function is constructed by comprehensively considering uncertainties, spectral spatial gradient, and UAV flight costs:

[0049] ,

[0050] in, The score represents the candidate sampling location. This indicates uncertainty at that location. This indicates the degree of spectral variation near that location. Indicates the cost of drone flight, , , These are the weighting coefficients.

[0051] Step 603: Select the highest-scoring location from the candidate regions near the drone's current position as the next sampling location: .in, This indicates the candidate sampling area near the current location of the drone. Indicates the selected next sampling position.

[0052] Step 7 above specifically includes: the UAV flying to the next sampling location and collecting new spectrum observation data, and incorporating the new data into the sparse spectrum observation dataset. The data is then input into the large spectrum situation model again for reconstruction and trajectory update, thus forming a closed-loop process of UAV autonomous sampling and three-dimensional spectrum situation construction.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. This invention deeply integrates UAV autonomous trajectory planning with 3D spectrum situation reconstruction, forming a closed-loop feedback mechanism. Compared with existing technologies where sampling and reconstruction are independent and the UAV flies along a fixed trajectory or preset path, this invention can dynamically guide the UAV to prioritize supplementary sampling in areas with the highest information value based on the uncertainty distribution in the current reconstruction results. This mechanism maximizes data utilization efficiency under a limited sampling budget, avoids resource waste caused by blind sampling and repeated sampling, and achieves higher reconstruction accuracy under the same number of sampling points.

[0055] 2. This invention constructs a dedicated Transformer large-scale model for spectrum reconstruction tasks in low-altitude 3D scenes. This model, through structures such as patch embedding, hierarchical encoding, and skip connections, can simultaneously capture local details and long-range spatial dependencies in the spectrum distribution. Compared with traditional methods such as Kriging interpolation and inverse distance weighted interpolation, this invention does not rely on the assumption of local smoothness and has stronger expressive power for signal abrupt changes and complex boundaries. Compared with conventional deep learning models (such as U-Net and autoencoders), this invention, through window self-attention mechanism and trajectory-aware encoding, can better adapt to sparse and irregular UAV sampling patterns. Simulation results show that even with a sampling ratio as low as 1%, this invention can still recover the main spectrum distribution characteristics of the target region.

[0056] 3. Building upon data-driven reconstruction, this invention further introduces a physical perception correction mechanism based on building occlusion relationships. By constructing a line-of-sight / non-line-of-sight discrimination map and compensating for unobserved non-line-of-sight areas using prediction biases in observed non-line-of-sight areas, it effectively overcomes the inherent defects of pure deep learning models, such as oversmoothing and prediction bias in building-occluded areas. This mechanism makes the reconstruction results more consistent with the physical laws of low-altitude electromagnetic propagation, making it particularly suitable for application scenarios with significant occlusion, such as urban low-altitude areas, industrial park low-altitude areas, and densely built-up areas, significantly improving reconstruction stability and engineering applicability in complex environments.

[0057] 4. This invention, through a strategy of merging the height and channel dimensions, transforms the 3D reconstruction problem into a format that can be efficiently processed by a 2D Transformer while preserving 3D spatial information, avoiding the high computational overhead of directly using 3D convolution. Simultaneously, it fuses the sampling spectrum, sampling mask, building distribution, and height information into a multimodal input tensor, enabling the model to simultaneously perceive observation data, sampling distribution, and environmental constraints within a unified framework, providing a sufficient information foundation for high-quality reconstruction.

[0058] 5. This invention, through the synergistic effect of large model data-driven reconstruction, physical perception correction, and uncertainty-aware trajectory planning, maintains high accuracy in spectral situation construction and good reconstruction stability even under extremely low sampling ratios (0.5%~3%). Simulation results show that, under different sampling ratios, building densities, and signal-to-noise ratios, the root mean square error of this invention is significantly lower than that of traditional interpolation methods, conventional deep learning methods, and random flight sampling schemes, with minimal performance fluctuations. This characteristic enables this invention to complete high-precision low-altitude three-dimensional spectral situation construction tasks with low UAV flight energy consumption and short sampling time in practical deployments. Attached Figure Description

[0059] Figure 1This is a flowchart of the method of the present invention; Figure 2 This is a large model framework diagram of the present invention; Figure 3 This is a three-dimensional spectral situation reconstruction effect diagram using the present invention during the pre-training stage; Figure 4 This is a diagram illustrating the effect of fine-tuning and reconstructing the system using this invention in a complex low-altitude environment. Figure 5 This is a comparison chart of the reconstruction results of the present invention with those of random flight mode and various benchmark reconstruction methods; Figure 6 This is a comparison chart of the root mean square error changes of various methods under different sampling ratios. Detailed Implementation

[0060] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, the present invention provides a three-dimensional spectrum situation construction method based on UAV autonomous planning, comprising:

[0062] Step 1: Construct a three-dimensional spatial model of the low-altitude target area and divide the target area into grids to obtain a three-dimensional discrete spatial representation for spectrum situation construction;

[0063] Step 2: The UAV performs spectrum sampling according to the initial trajectory within the low-altitude target area to obtain the received signal strength at different spatial locations, forming sparse spectrum observation data;

[0064] Step 3: Based on the sparse spectrum observation data, sampling locations, building distribution information, and height information, construct the input data for the large spectrum situation model;

[0065] Step 4: Using a large-scale spectral situation model based on the Transformer structure, feature extraction and three-dimensional spectral situation reconstruction are performed on the input data to obtain the initial estimation results of the three-dimensional spectral situation of the target area;

[0066] Step 5: Based on the building occlusion relationship and the observed spectrum data, perform physical perception correction on the initial estimation result of the three-dimensional spectrum situation to obtain a three-dimensional spectrum situation map with enhanced physical consistency;

[0067] Step 6: Based on the uncertainty of the three-dimensional spectrum situation map, combined with the degree of spectrum spatial change and the flight cost of the UAV, determine the next sampling position and update the UAV sampling trajectory;

[0068] Step 7: The UAV continues to collect spectrum data according to the updated sampling trajectory, and inputs the newly sampled data into the spectrum situation large model for iterative reconstruction until the preset sampling budget or reconstruction accuracy requirements are met, and outputs a complete three-dimensional spectrum situation map.

[0069] To optimize the above technical solution, the specific measures also include:

[0070] (1) Acquisition of low-altitude three-dimensional spectrum data

[0071] Without loss of generality, this invention considers a typical low-altitude three-dimensional communication scenario and a typical transmission model, namely, assuming that there are S transmitting sources operating at a specific frequency f within the range of interest, which can be specifically expressed as:

[0072]

[0073] in, Indicates the received signal strength. Indicates the number of launch sources. Indicates the first Each transmitter is at frequency The transmit power at the following levels Indicates the corresponding channel gain. This represents noise and interference. In low-altitude network scenarios, the channel gain is not only affected by propagation distance, but also closely related to factors such as building obstruction, non-line-of-sight propagation, shadow fading, and multipath propagation.

[0074] In the process of constructing a three-dimensional spectral situation, this invention represents the low-altitude target region as a three-dimensional spatial region: To facilitate modeling and calculation, this three-dimensional spatial region is discretized into the following parts: A three-dimensional mesh, in which , , These represent the number of grid cells in the horizontal, vertical, and vertical directions of the target region, respectively; let the actual size of the target region be... , , Then the grid spacing in each direction is:

[0075] ,

[0076] in, , , Through the discretization process described above, the electromagnetic field strength distribution in continuous low-altitude space is converted into a three-dimensional spectral situation tensor that can be processed by the model.

[0077] Furthermore, the UAV sequentially visits several sampling locations within the target area along its flight path, and collects the received signal strength at each location. The sparse spectrum observation dataset obtained by the UAV is represented as follows: in, This represents a sparse spectral observation dataset. This indicates the number of sampled points. Due to limitations in UAV endurance, flight energy consumption, and sampling frequency, the actual number of observation points obtainable is usually far less than the number of points in a complete 3D grid. Therefore, a large model is needed to recover a complete 3D spectral situation map from sparse observation data.

[0078] (2) Model input construction

[0079] In this embodiment, in order to enable the large spectrum situation model to simultaneously perceive sampling results and environmental information, the present invention uses the sampling spectrum map, sampling mask map, building distribution map and height information map as model inputs.

[0080] Specifically, the model input tensor is represented as: in, The number of input channels is represented. The sampled spectrum is used to represent the received signal strength at the locations already acquired by the UAV; the sampled mask map is used to distinguish between observed and unobserved locations; the building distribution map is used to characterize obstructing structures in the low-altitude environment; and the height information map is used to preserve the spatial differences between different height layers. To reduce the computational complexity of direct processing of 3D data, this invention folds the height dimension into the channel dimension, resulting in the transformed input tensor: Through the above processing, the model can retain three-dimensional height information while employing a two-dimensional feature extraction structure, thus balancing computational efficiency and low-altitude three-dimensional spectrum modeling capabilities.

[0081] (3) Construction of a large-scale three-dimensional spectrum situation model

[0082] In this embodiment, step 4 constructs a large-scale spectral situation model based on the Transformer structure. This large-scale model is centered on a hierarchical Transformer structure and mainly includes a Patch embedding module, a hierarchical encoder, a trajectory-aware encoder module, a hierarchical decoder, and a jump connection module.

[0083] First, the Patch embedding module will take the input tensor The image is divided into several non-overlapping local patches and mapped to a high-dimensional feature space. Let the size of each patch be... The number of patches is: After linear mapping, the input feature sequence is obtained: 1. Among them, Indicates the number of patches. The embedding vector dimension is represented. Subsequently, a hierarchical encoder performs multi-scale spatial feature extraction on the input features. The encoder models local spectral correlations through a windowed multi-head self-attention mechanism and expands the receptive field through progressive downsampling, thereby capturing long-range spatial dependencies in low-altitude spectral situations. Simultaneously, the trajectory-aware encoding module maps the UAV sampling trajectory and sampling mask information to the feature space and fuses it with spectral features, enabling the model to adapt to sparse and irregular UAV sampling patterns.

[0084] During the decoding phase, the hierarchical decoder gradually recovers the spatial resolution through patch expansion and skip connections, and outputs an initial estimate of the three-dimensional spectral situation. Its reconstruction process can be represented as: in, This represents the reconstructed three-dimensional spectral situation map. Indicates by parameters A large-scale model representing the spectral situation. During model training, this invention primarily calculates the reconstruction error for unobserved regions, enabling the model to focus on learning how to recover the complete spectral distribution from sparse observation data. The loss function is expressed as:

[0085] ,

[0086] in, Represents the set of unobserved locations. Indicates position The actual spectral value at that location, Indicates position The reconstructed spectrum values ​​at the location. Through the above training method, the model can learn the spatial distribution law of low-altitude electromagnetic propagation under sparse sampling conditions, thereby improving the accuracy of the three-dimensional spectrum situation map reconstruction.

[0087] (4) Physical perception correction

[0088] In this embodiment, since building obstruction in low-altitude urban environments causes significant non-line-of-sight propagation and signal abrupt changes, relying solely on data-driven models can easily lead to over-smoothing or prediction bias in obstructed areas. Therefore, this invention further introduces a physical perception correction mechanism to compensate for the initial estimation results of the three-dimensional spectral situation.

[0089] First, based on the building distribution information and the estimated location of the emission source, it is determined whether the direct path from the emission source to each grid location is blocked, and a line-of-sight propagation discrimination map is constructed:

[0090] ,

[0091] in, This indicates that the location is within the line-of-sight propagation area. This indicates that the location is a non-line-of-sight occlusion area. For non-line-of-sight areas, the prediction bias is estimated based on the observed samples:

[0092] ,

[0093] in, This indicates the number of samples observed in non-line-of-sight regions. This represents the average prediction bias in the non-line-of-sight region. Finally, the average prediction bias is used to correct the non-line-of-sight region, resulting in a three-dimensional spectral situation map with enhanced physical consistency.

[0094] ,

[0095] in, This represents the spectral state value after physical perception correction.

[0096] (5) Uncertainty-aware UAV autonomous planning

[0097] In this embodiment, step 6 plans the next sampling location for the UAV based on the uncertainty of the reconstruction results. Since the initial observation data is relatively sparse, the model's prediction results in some areas still have uncertainty. Therefore, it is necessary to guide the UAV to prioritize areas with higher information content for supplementary sampling. Specifically, this invention performs multiple random forward propagations during the model inference stage to obtain multiple three-dimensional spectral situation prediction results, and constructs an uncertainty graph based on the variance between different prediction results. When determining the next sampling location, a sampling scoring function is constructed by comprehensively considering uncertainties, spectral spatial gradients, and UAV flight costs:

[0098] ,

[0099] in, The score represents the candidate sampling location. This indicates uncertainty at that location. This indicates the degree of spectral variation near that location. Indicates the cost of drone flight, , , These are the weighting coefficients.

[0100] The next sampling position is represented as: .in, This indicates the candidate sampling area near the current location of the drone. Indicates the selected next sampling position.

[0101] The UAV generates a feasible flight trajectory based on the next sampling location and continues to collect new spectrum data during flight. The newly sampled data is incorporated into the sparse spectrum observation dataset. The model input is then reconstructed, and three-dimensional spectral situation reconstruction, physical perception correction, and the next round of trajectory planning are performed. Through the above methods, this invention forms a closed-loop process of "sampling-reconstruction-correction-planning-resampled".

[0102] The effects of the present invention will be further explained below with reference to simulation experiments.

[0103] 1. Simulation conditions and parameter settings:

[0104] The simulation experiment of this invention is based on a low-altitude three-dimensional communication scenario. As shown in Table 1, the target area size is set to 1000×1000×100m. 3 The 3D grid resolution was set to 128×128×5, and the carrier frequency was set to 2.45GHz. One to three transmission sources were randomly set within the target area, with a transmission power range of 10dBm to 15dBm. The number of buildings was set to 0 to 3, and the sampling ratio was set to 0.5% to 10%.

[0105] Table 1

[0106] Area size <![CDATA[L x × L y × H z ]]> <![CDATA[1000 × 1000 × 100 m 3 ]]> Grid resolution <![CDATA[N x × N y × N z ]]> 128 × 128 × 5 carrier frequency f 2.45 GHz Number of launch sources S [1, 3] Transmit power <![CDATA[P t,s ]]> [10, 15] dBm Number of buildings <![CDATA[N b ]]> [0, 3] Sampling ratio r 0.5%～10%

[0107] As shown in Table 2, the large-scale spectral situation model used has 20 input channels, a patch size of 4×4, an embedding dimension of 192, and the encoder and decoder adopt a Transformer structure with a window size of 8×8. These settings effectively simulate the complex electromagnetic environment in low-altitude networks under the combined effects of multiple emission sources, building obstruction, and three-dimensional height variations.

[0108] Table 2

[0109] Number of input channels C 20 Patch size P 4 × 4 Embedded vector dimension D 192 Encoder / decoder depth <![CDATA[L enc / L dec ]]> (2, 2, 8, 2) Number of attention heads <![CDATA[H l ]]> (6, 12, 24, 48) Window size W 8 × 8

[0110] 2. Simulation content:

[0111] Figure 3 This demonstrates the effectiveness of the present invention in reconstructing the three-dimensional spectral situation during the pre-training phase. Even with a sampling rate of only 1%, the model is still able to recover the main spectral distribution characteristics within the target region relatively well, indicating that the large spectral situation model can learn fundamental electromagnetic propagation laws from a very small amount of observational data.

[0112] Figure 4 This paper demonstrates the fine-tuning reconstruction effect of the present invention in complex low-altitude environments. Compared with the pre-training stage, the fine-tuning stage further considers complex factors such as building obstruction, scattering, and multipath propagation, significantly increasing the difficulty of spectral situation reconstruction. Simulation results show that the present invention can still achieve relatively stable reconstruction results under the condition of a sampling ratio of 3%, indicating that the proposed method has good adaptability to complex low-altitude propagation environments.

[0113] Figure 5 The results show a comparison between this invention and random flight methods, as well as various benchmark reconstruction methods. It can be seen that, compared to random flight, this invention can guide the UAV to prioritize collecting samples with higher information value, thus obtaining more accurate reconstruction results. Compared to methods such as U-Net, Vision Transformer, inverse distance weighted interpolation, K-nearest neighbors, and Kriging interpolation, this invention exhibits better boundary recovery and spatial detail preservation capabilities in regions with signal abrupt changes and building occlusion.

[0114] Figure 6 The paper further demonstrates the variation of root mean square error (RMSE) for each method under different sampling ratios. As the sampling ratio changes, this invention consistently maintains low reconstruction error and minimal performance fluctuations, indicating that the proposed uncertainty-aware sampling strategy and physical-aware correction mechanism can effectively improve the stability and robustness of 3D spectral situation construction.

[0115] Based on the simulation results and analysis above, the 3D spectrum situation construction method based on UAV autonomous planning proposed in this invention can achieve high-precision 3D spectrum situation reconstruction under low sampling ratio conditions. This method not only improves the spectrum sampling efficiency of UAVs but also enhances the model's adaptability in environments with building obstruction, non-line-of-sight propagation, and complex low-altitude environments. It is suitable for applications such as low-altitude network spectrum sensing, interference analysis, wireless network planning, and dynamic spectrum resource management. The above are merely preferred embodiments of this invention, and the scope of protection of this invention is not limited to the above embodiments. All technical solutions falling within the scope of this invention's concept are within the scope of protection of this invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this invention should be considered within the scope of protection of this invention.

Claims

1. A method for constructing a three-dimensional spectrum situation based on autonomous planning by unmanned aerial vehicles (UAVs), characterized in that, include: Step 1: Construct a three-dimensional spatial model of the low-altitude target area and divide the target area into grids to obtain a three-dimensional discrete spatial representation for spectrum situation construction; Step 2: The UAV performs spectrum sampling according to the initial trajectory within the low-altitude target area to obtain the received signal strength at different spatial locations, forming sparse spectrum observation data; Step 3: Based on the sparse spectrum observation data, sampling locations, building distribution information, and height information, construct the input data for the large spectrum situation model; Step 4: Using a large-scale spectral situation model based on the Transformer structure, feature extraction and three-dimensional spectral situation reconstruction are performed on the input data to obtain the initial estimation results of the three-dimensional spectral situation of the target area; Step 5: Based on the building occlusion relationship and the observed spectrum data, perform physical perception correction on the initial estimation result of the three-dimensional spectrum situation to obtain a three-dimensional spectrum situation map with enhanced physical consistency; Step 6: Based on the uncertainty of the three-dimensional spectrum situation map, combined with the degree of spectrum spatial change and the flight cost of the UAV, determine the next sampling position and update the UAV sampling trajectory; Step 7: The UAV continues to collect spectrum data according to the updated sampling trajectory, and inputs the newly sampled data into the spectrum situation large model for iterative reconstruction until the preset sampling budget or reconstruction accuracy requirements are met, and outputs a complete three-dimensional spectrum situation map.

2. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 1 specifically includes: Representing the low-altitude target region as a three-dimensional spatial region: And divide it into: A three-dimensional mesh, in which , , These represent the number of grid cells in the horizontal, vertical, and vertical directions of the target region, respectively; let the actual size of the target region be... , , Then the grid spacing in each direction is: ， in, , , .

3. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 2 specifically includes: Step 201: The drone's spatial location and frequency The received signal strength acquired below is expressed as: ， in, Indicates the received signal strength. Indicates the number of launch sources. Indicates the first Each transmitter is at frequency The transmit power at the following levels Indicates the corresponding channel gain. Indicates noise and interference terms. Step 202: The UAV collects the received signal strength at several locations along its flight path to form a sparse spectrum observation dataset. ,in, This represents a sparse spectral observation dataset. This indicates the number of sampled points.

4. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 3 specifically includes: Step 301: Construct a multimodal input tensor by using the sampled spectrogram, sampling mask, building distribution information, and height information as model inputs: in, Indicates the number of input channels. Step 302: To facilitate model processing, the height dimension and channel dimension are merged to obtain the transformed input tensor, resulting in: .

5. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 4 specifically includes: Step 401: Use a large spectrum situation model based on the Transformer structure as the reconstruction network. The large spectrum situation model includes a Patch embedding module, a hierarchical encoder, a trajectory-aware encoder, a hierarchical decoder, and a jump connection module. Step 402: The input tensor is divided into several local blocks by the Patch embedding module and mapped to a high-dimensional feature space; spatial correlation features at different scales are extracted by the hierarchical encoder; the UAV sampling trajectory and sampling mask information are fused by the trajectory-aware encoding module; the spatial resolution is restored by the hierarchical decoder, and the three-dimensional spectral situation estimation result is output. Step 403: The three-dimensional spectral situation reconstruction process is represented as follows: ,in, This represents the reconstructed three-dimensional spectral situation map. Indicates by parameters A large model representing the spectral situation. Step 404: During model training, the reconstruction error is calculated for the unobserved region, and the loss function is expressed as: ， in, Represents the set of unobserved locations. Indicates position The actual spectral value at that location, Indicates position The reconstructed spectral value at that location.

6. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 5 specifically includes: Step 501: Based on the building distribution information and the estimated location of the emission source, determine whether there is occlusion between the emission source and each grid location, and construct a line-of-sight propagation discrimination map: ， in, This indicates that the location is within the line-of-sight propagation area. This indicates that the location is a non-line-of-sight occlusion area. For non-line-of-sight areas, the prediction bias is estimated based on the observed samples: ， in, This indicates the number of samples observed in non-line-of-sight regions. This represents the average prediction bias in non-line-of-sight areas. Step 502: Correct the non-line-of-sight region using the average prediction bias to obtain a three-dimensional spectral situation map with enhanced physical consistency: ， in, This represents the spectral state value after physical perception correction.

7. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 6 specifically includes: Step 601: Perform multiple random forward propagations during the model inference phase to obtain multiple three-dimensional spectral situation prediction results, and calculate the uncertainty graph based on the variance between different prediction results. , Step 602: When determining the next sampling location, a sampling scoring function is constructed by comprehensively considering uncertainties, spectral spatial gradient, and UAV flight costs: ， in, The score represents the candidate sampling location. This indicates uncertainty at that location. This indicates the degree of spectral variation near that location. Indicates the cost of drone flight, , , These are the weighting coefficients; Step 603: Select the highest-scoring location from the candidate regions near the drone's current position as the next sampling location: ,in, This indicates the candidate sampling area near the current location of the drone. Indicates the selected next sampling position.

8. The method for constructing a three-dimensional spectrum situation based on UAV autonomous planning according to claim 1, characterized in that, Step 7 specifically includes: The drone flies to the next sampling location and collects new spectrum observation data, which is then incorporated into the sparse spectrum observation dataset. The data is then input into the large spectrum situation model again for reconstruction and trajectory update, thus forming a closed-loop process of UAV autonomous sampling and three-dimensional spectrum situation construction.

Images