A method for synthesizing three-dimensional radio map data by fusing measured data

Abstract

This invention discloses a method for synthesizing three-dimensional radio map data by fusing measured data, comprising the following steps: Step S1: Acquire geographic environment data and transmitter configuration data of the target area, and convert them into tensor form; Step S2: Obtain preliminary three-dimensional map prediction results; Step S3: Construct a multidimensional generalized distance mapping function; Step S4: Based on the multidimensional generalized distance mapping function, construct a generalized parameterized target model representing the logarithmic scale received signal strength at any location in space; Step S5: Optimize the coefficient combination of the target model based on measured data, and aggregate to form a coefficient matrix, representing the feasible region of the coefficients; Step S6: Resample within the feasible region of the coefficients and synthesize three-dimensional data to obtain three-dimensional radio map data samples. This invention features low cost, high efficiency, and high physical fidelity.

Description

Technical Field

[0001] This invention relates to the intersection of wireless communication and artificial intelligence, specifically to a method for synthesizing three-dimensional radio map data that integrates measured data. Background Technology

[0002] In the 3D Radio Map data synthesis scheme for sixth-generation (6G) communication and low-altitude networks (such as UAV communication scenarios), the radio map, as an important tool for characterizing the distribution of spatial radio resources (such as received signal strength, path loss, etc.), is the foundation for realizing network resource optimization, interference management, and UAV trajectory planning.

[0003] In recent years, deep learning (DL) technology has demonstrated its potential to surpass traditional ray-tracing methods in radio map estimation (RME) tasks due to its extremely high inference efficiency. However, most existing deep learning-based radio map construction methods are limited to near-ground two-dimensional (2D) scenes, making it difficult to meet the practical needs of low-altitude economic and 3D spatial network coverage. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for synthesizing three-dimensional radio map data that integrates measured data. This method organically integrates two-dimensional simulation data with a very small amount of measured three-dimensional data, and features low cost, high efficiency and high physical fidelity.

[0005] The objective of this invention is achieved through the following technical solution: a method for synthesizing three-dimensional radio map data by fusing measured data, comprising the following steps:

[0006] Step S1: Obtain the geographic environment data and transmitter configuration data of the target area, and convert them into tensor form;

[0007] Step S2: Use a pre-trained two-dimensional neural network to perform layer-by-layer inference to obtain preliminary three-dimensional map prediction results;

[0008] Step S3: Construct a multidimensional generalized range mapping function, including a composite function for three-dimensional large-scale fading and a two-dimensional horizontal range mapping function for non-ideal antenna radiation patterns;

[0009] Step S4: Based on the multidimensional generalized distance mapping function, construct a representation space for any location. A generalized parameterized target model for the received signal strength at a logarithmic scale;

[0010] Step S5: Optimize the coefficient combination of the target model based on the measured data, and aggregate them to form a coefficient matrix, which represents the feasible region of the coefficients;

[0011] Step S6 involves resampling within the feasible region of the coefficients and synthesizing the three-dimensional data to obtain a three-dimensional radio map data sample.

[0012] The beneficial effects of this invention are: 1. It breaks through the bottleneck of scarce three-dimensional radio map data and realizes large-scale data synthesis at extremely low cost: Compared with the high computation and time cost of traditional three-dimensional ray tracing methods that often take several hours, the method of this invention only needs to perform a single two-dimensional model inference and basic mathematical model calculation, and the generation efficiency is improved by orders of magnitude (reaching linear time complexity), which is extremely suitable for the rapid generation of massive amounts of data.

[0013] 2. High physical fidelity, restoring the unique electromagnetic propagation characteristics of three-dimensional space: By innovatively introducing generalization functions about three-dimensional spatial distance and two-dimensional horizontal distance as empirical correction terms, the synthesized data can accurately reflect the three-dimensional channel fading and antenna polarization effects in the real physical world, solving the problem of spatial signal distortion caused by existing two-dimensional dimensionality reduction schemes.

[0014] 3. Possesses extremely strong data diversity and scenario generalization capabilities: This solution requires only a very small amount (even a single flight acquisition) of real measurement data to complete model calibration. By extracting the feasible domain of coefficients and resampling within the domain, it can flexibly synthesize data samples covering various complex signal propagation characteristics, laying a high-quality data foundation for subsequent training of highly robust 3D deep learning networks (such as 3D U-Net or 3D-GAN). Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention;

[0016] Figure 2 This is a schematic diagram illustrating the process of generating a large-scale 3D synthetic dataset. Detailed Implementation

[0017] 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.

[0018] This invention organically integrates two-dimensional simulation data (prediction results based on pre-trained two-dimensional models) with a very small amount of three-dimensional measured data (such as data collected by UAVs) to construct a generalized parameterized target model characterizing three-dimensional spatial fading and antenna polarization characteristics. Based on the obtained two-dimensional prediction map, this scheme is not limited to fixed empirical formulas, but innovatively introduces a generalization function for three-dimensional spatial distance. And the generalization function for two-dimensional horizontal distance As an empirical correction term, this generalized parameterized model perfectly compensates for the technical deficiency of two-dimensional models in being unable to characterize three-dimensional channel fading and antenna polarization effects. Low-cost parameter calibration and ultra-large-scale data augmentation based on a small amount of real measurement data: using a very small amount of real measurement data, the optimal set of coefficients for the above parameterized model is determined through optimization algorithms (such as least squares regression); then, the feasible range of these coefficients (feasible region) is extracted, and through flexible resampling within the feasible region, a large-scale and high-quality synthetic dataset of three-dimensional radio maps covering a vast number of different signal propagation characteristics is efficiently generated with extremely low computational complexity (linear complexity). This scheme breaks the technical deadlock in the field of three-dimensional radio maps, which is characterized by "difficulty in obtaining real data and slow simulation generation."

[0019] Specifically, such as Figure 1 As shown, a method for synthesizing three-dimensional radio map data by fusing measured data includes the following steps:

[0020] Step S1: Obtain the geographic environment data and transmitter configuration data of the target area, and convert them into tensor form;

[0021] Specifically, it includes:

[0022] Geographic environment information construction: Extract geographic environment data of the target area, i.e., 3D building maps, convert them into voxel meshes with uniform spatial resolution, and generate 3D binary occupancy tensors. Where 1 indicates that the voxel mesh is occupied by buildings, 0 indicates free space, and W, D, and H represent the width, depth, and height of the target area, respectively;

[0023] Transmitter information construction: Obtain transmitter configuration data, including A set of three-dimensional coordinates of each 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.

[0024] Step S2: Use a pre-trained two-dimensional neural network to perform layer-by-layer inference to obtain preliminary three-dimensional map prediction results;

[0025] Specifically, it includes:

[0026] 1. Dimensionality reduction slicing: For a given height With respect to the transmitter tensor along Projection operation on axis And extract height The corresponding two-dimensional building matrix Combined into a two-dimensional feature vector .

[0027] 2. 2D Inference and 3D Reconstruction: Input pre-trained two-dimensional model In the middle, the height is calculated. Corresponding two-dimensional radio map prediction value All heights The two-dimensional predicted map is stitched together along the height dimension to obtain a preliminary predicted three-dimensional radio map. .

[0028] In the embodiments of this application, the pre-trained two-dimensional model can be a neural network model, using several sets of known... , Pre-training;

[0029] Step S3: Construct a multidimensional generalized range mapping function, including a composite function for three-dimensional large-scale fading and a two-dimensional horizontal range mapping function for non-ideal antenna radiation patterns;

[0030] 1. Transformation design for large-scale 3D fading ( ):

[0031] Composite functions are used to characterize three-dimensional path loss. .in, The power transform distribution parameters, This is used to characterize the linear loss coefficient of the medium absorption. Specifically, when... and At that time, the function mathematically degenerates into the traditional form. By introducing and This generalization function not only covers the logarithmic decay law, but can also adaptively fit the steeper electromagnetic wave energy dissipation curve under non-line-of-sight (NLOS) conditions. This represents the three-dimensional Euclidean distance between the receiver and the transmitter.

[0032] 2. Transformation design for non-ideal antenna radiation patterns ( ):

[0033] Because the polarization gain of a real antenna is not only related to the sine of the relative elevation angle ( The invention relates to the two-dimensional horizontal distance mapping function and is also affected by the antenna's mechanical structure and beamforming. Designed as a truncated Taylor series expansion of the number field. .in, The order of the polynomial expansion is set to the preset order (take...). or To balance computational complexity. Weights are fitted for each order of polarization. and At that time, the function degenerates into the basis for adapting an ideal dipole antenna. Model; This represents the two-dimensional relative distance between the receiving point and the transmitter on the horizontal projection plane. When higher-order terms are introduced (…), When the model has a strong envelope fitting capability, it can adaptively learn and approximate the complex pattern distortion produced by any non-ideal antenna at different horizontal distances, which greatly expands the application scope of synthetic data.

[0034] Step S4: Based on the multidimensional generalized distance mapping function, construct a representation space for any location. A generalized parameterized target model for the received signal strength at a logarithmic scale;

[0035] To overcome the limitations of two-dimensional models in characterizing three-dimensional channel fading and antenna polarization effects, a model is constructed to represent arbitrary locations in space. Generalized parameterized target model for received signal strength at logarithmic scale :

[0036]

[0037] In the formula:

[0038] The transmitter location; The preliminary predicted 3D radio map obtained in step 2 is located in... The value at that location.

[0039] For receiving point With the transmitter position The three-dimensional Euclidean distance between them; For receiving point With the transmitter position Two-dimensional relative distance on the horizontal (XY) projection plane;

[0040] and For the function in step S3, the two-dimensional network baseline features are... It is integrated with the physical decay characteristics of three-dimensional space.

[0041] The combination of coefficients for the target model to be determined.

[0042] Step S5: Optimize the coefficient combination of the target model based on the measured data, and aggregate them to form a coefficient matrix, which represents the feasible region of the coefficients;

[0043] Using samples containing real measurements (such as actual received signal strength collected by a drone platform) The model in step S4 is calibrated. Specifically, this includes:

[0044] 1. Error minimization optimization: minimizing the difference between the calculated and measured values ​​of the parameterized model. Taking the difference between them as the objective, construct error minimization problems such as least squares regression:

[0045]

[0046] Solving this problem yields a set of optimal coefficients corresponding to a single measurement. . Includes multiple actual received signal strengths Each Corresponding to a receiving point ;

[0047] 2. Feasible domain construction: using A three-dimensional measured data sample containing different geographical environments and transmitter distributions was used to repeat the above optimization process to obtain... Group the optimal coefficient vectors and aggregate them to form a coefficient matrix. This matrix defines the feasible range (feasible region) of each coefficient under reasonable physical constraints in a real physical environment (e.g., wait) .

[0048] Step S6 involves resampling within the feasible region of the coefficients and synthesizing the three-dimensional data to obtain a three-dimensional radio map data sample.

[0049] The computing device calculates the coefficients within the feasible region determined in step 5. Perform independent resampling or joint oversampling to generate Group( Combinations of derived coefficients that satisfy the physical reality distribution For a specific geographical environment and a specific transmitter distribution, any combination of derived coefficients can be selected. By substituting the parameterized target model, the signal intensity of all voxel points in space can be directly calculated, and a high-quality 3D radio map data sample covering extremely rich signal propagation characteristics can be synthesized. :

[0050] Synthesized 3D radio map data samples It includes all receiving points. The actual received signal strength at the location;

[0051] For synthesizing three-dimensional radio map data samples ,exist The location will be updated to match the actual received signal strength with the transmit power. ;

[0052] In horizontal distance and The location is directly above or below the transmitter point. At this time, the signal strength is set to 0 or the near-field strength. The near-field strength is set to a fixed constant value, which is the noise limit in this environment.

[0053] Repeating the above process allows for the efficient synthesis of large-scale 3D synthetic datasets (Radio3DMix) needed for subsequent training (such as 3D-GAN models). Figure 2 As shown.

[0054] In the embodiments of this application, a comparison with existing synthetic datasets is shown in Table 1 below. It can be seen that the constructed dataset has the following advantages compared with other datasets: it can cover large-scale 3D high-precision data and supports multiple transmitters.

[0055] Table 1 Comparison of existing synthetic datasets

[0056]

[0057] 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 synthesizing three-dimensional radio map data by fusing measured data, characterized in that: Includes the following steps: Step S1: Obtain the geographic environment data and transmitter configuration data of the target area, and convert them into tensor form; Step S1 includes the following sub-steps: S101: Geographic Environment Information Construction: Extract geographic environment data of the target area, i.e., 3D building maps, convert them into voxel meshes with uniform spatial resolution, and generate 3D binary occupancy tensors. Where 1 indicates that the voxel mesh is occupied by buildings, 0 indicates free space, and W, D, and H represent the width, depth, and height of the target area, respectively; S102: Transmitter Information Construction: Obtain transmitter configuration data, including A set of three-dimensional coordinates of each transmitter and its transmission power Construct the transmitter information tensor If a transmitter exists at a voxel grid location, then the element value at that voxel grid location is... Otherwise, it is 0; Step S2: Use a pre-trained two-dimensional neural network to perform layer-by-layer inference to obtain preliminary three-dimensional map prediction results; Step S2 includes: S201. Dimensional Reduction Slicing: For a given height With respect to the transmitter tensor along Projection operation on axis and for binary occupied tensors Extraction height The corresponding two-dimensional building matrix Combined into a two-dimensional feature vector ; S202.2D Inference and 3D Reassembly: Input pre-trained 2D model middle, Calculate the height for the model parameters. Corresponding two-dimensional radio map prediction value ; All heights The two-dimensional predicted map is stitched together along the height dimension to obtain a preliminary predicted three-dimensional radio map. ; Step S3: Construct a multidimensional generalized range mapping function, including a composite function for three-dimensional large-scale fading and a two-dimensional horizontal range mapping function for non-ideal antenna radiation patterns; Step S3 includes: Constructing a composite function for large-scale fading in three dimensions Characterized by three-dimensional path loss, denoted as: ； in, The power transform distribution parameters, The linear loss coefficient characterizing the absorption of the medium; This represents the three-dimensional Euclidean distance between the receiver and the transmitter. Mapping function of two-dimensional horizontal distance The design is a truncated Taylor series expansion of the logarithmic field, denoted as: ； in, The order of the polynomial expansion is preset. The polarization fitting weights are for the m-th order. This represents the two-dimensional relative distance between the receiving point and the transmitter on the horizontal projection plane. Step S4: Based on the multidimensional generalized distance mapping function, construct a representation space for any location. A generalized parameterized target model for the received signal strength at a logarithmic scale; Step S4 includes: to compensate for the inability of two-dimensional models to characterize three-dimensional channel fading and antenna polarization effects, constructing a model representing arbitrary locations in space. Generalized parameterized target model for received signal strength at logarithmic scale : ； In the formula: The transmitter location; To make preliminary predictions of the location of the three-dimensional radio map The value at; For receiving point With the transmitter position The three-dimensional Euclidean distance between them; For receiving point With the transmitter position Two-dimensional relative distance on the horizontal projection plane; and Used to define the baseline features of a two-dimensional network Integrate with three-dimensional spatial physical decay characteristics; The combination of coefficients for the target model to be determined; Step S5: Optimize the coefficient combination of the target model based on the measured data, and aggregate them to form a coefficient matrix, which represents the feasible region of the coefficients; Step S5 includes: S501. Error Minimization Optimization: Minimizing the difference between the calculated and measured values ​​of the parameterized model. Taking the difference between them as the objective, we construct an error minimization problem based on least squares regression: ； Solving for the optimal set of coefficients corresponding to a single measurement data yields a set of coefficients. The measured value The actual received signal strength collected by the drone platform; Includes multiple actual received signal strengths Each Corresponding to a receiving point ; S502. Feasibility Domain Construction: Using A three-dimensional measured data sample containing different geographical environments and transmitter distributions is obtained by repeating the optimization process in step S501. The optimal coefficient vectors are aggregated to form a coefficient matrix. , is used to define the feasible region of each coefficient under reasonable physical constraints in a real physical environment; Step S6 involves resampling within the feasible region of the coefficients and synthesizing the three-dimensional data to obtain a three-dimensional radio map data sample.

2. The method for synthesizing three-dimensional radio map data by fusing measured data according to claim 1, characterized in that: Step S6 includes: Within the defined feasible region of coefficients, for the coefficients Perform independent resampling or joint oversampling to generate The group satisfies the derived coefficient combination of the physical reality distribution , ; Given a specific geographical environment and transmitter distribution, any combination of derived coefficients can be selected. By substituting the parameterized target model, the signal intensity of all voxel points in space is directly calculated, and a three-dimensional radio map data sample is synthesized. It includes all receiving points. The actual received signal strength at the location; For synthesizing three-dimensional radio map data samples ,exist The location will be updated to match the actual received signal strength with the transmit power. ; In horizontal distance and The location is directly above or below the transmitter point. At this time, the signal strength is set to 0 or the near-field strength. The near-field strength is set to a fixed constant value, which is the noise limit in this environment.

Images