Rapid wireless channel modeling method and system based on voxelization radiation field

Abstract

The invention provides a rapid wireless channel modeling method and system based on a voxelization radiation field. The method comprises the steps that a space to be modeled is divided into a three-dimensional voxel grid, and each voxel stores volume density and feature vectors; sampling along rays from the receiver Rx to the target direction to obtain a plurality of sampling points; inputting the transmitter Tx position and the feature vector of the sampling point into a deformation network, and outputting a corrected feature value; predicting the intensity of a signal emitted to the Rx by the sampling point by combining the corrected characteristic value and the ray direction information through a radiation network; calculating accumulated receiving signals of all sampling points on each ray to Rx based on a volume rendering model; a model is trained by using an error between a spatial spectrum predicted value and a real observation value and combining a weighted sum of a background entropy loss function as an optimization target. Through combination of space voxel modeling and a dual-network decoupling structure, channel space spectrum distribution at any transmitter position can be efficiently fitted, and the method is excellent in performance in a multipath complex environment.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a method and system for fast wireless channel modeling based on voxelized radiation fields. Background Art

[0002] In recent years, with the rapid development of sixth-generation mobile communication systems (6G), the communication frequency band has continued to expand to higher frequencies, and application scenarios have become increasingly complex and diverse. Traditional wireless channel modeling methods are facing severe challenges. As a fundamental step in the design, deployment, optimization, and evaluation of communication systems, the accuracy and efficiency of wireless channel modeling have a direct impact on overall system performance. Therefore, there is an urgent need to develop a new generation of modeling methods with higher accuracy, higher efficiency, and good generalization capabilities.

[0003] Existing wireless channel modeling methods can be categorized into three main types: deterministic modeling, statistical modeling, and hybrid modeling. Deterministic modeling methods typically rely heavily on environmental modeling, making them difficult to adapt to dynamic and complex scenarios. They also suffer from high computational overhead and are unsuitable for systems with high real-time requirements. Statistical modeling methods lack detailed descriptions of specific scenarios and cannot provide positional information such as angular distribution, making it difficult to achieve precise, task-oriented modeling. Hybrid modeling combines deterministic and statistical modeling, but this approach still relies on scene geometry information, presenting challenges in practical deployment.

[0004] In recent years, with the advancement of computer vision, Neural Radiance Fields (NeRF) has garnered widespread attention as an emerging 3D reconstruction technique. NeRF implicitly represents 3D scenes using a multi-layer perceptron (MLP), enabling high-quality view synthesis with only sparse images. Inspired by this approach, previous studies have attempted to extend NeRF for wireless channel modeling, modeling the scene as a wireless radiation field to generate a spatial angular power spectrum. However, due to its reliance on deep MLP networks and dense sampling, NeRF suffers from long training times and slow inference speed, making it difficult to meet the demands of real-time applications. Furthermore, some methods require auxiliary equipment such as LiDAR to acquire environmental point cloud data, increasing system complexity and deployment costs. In summary, while existing wireless channel modeling methods based on NeRF have achieved breakthroughs in accuracy, they still suffer from limitations in modeling speed, data dependency, and deployment flexibility. Therefore, a new channel modeling framework is urgently needed that can significantly improve modeling efficiency while maintaining high accuracy, reduce reliance on training data and external sensors, and meet the urgent need for efficient channel modeling in future 6G wireless systems.

[0005] Chinese patent publication CN118826933A discloses a wireless channel response prediction method, apparatus, device, and storage medium. The method comprises: receiving input prediction point data, the prediction point data including a predicted frequency and prediction point coordinates; acquiring a signal ray set with the prediction point coordinates as the center coordinates, encoding the signal ray set to obtain a signal direction code, wherein the signal ray set includes at least two virtual signal rays; uniformly arranging sampling voxel points along the virtual signal rays using a preset sampling distance and a preset sampling interval, and encoding the sampling voxel points to obtain a voxel position code; inputting the predicted frequency, voxel position code, and signal direction code into a preset neural network, outputting voxel density characteristics, voxel signal amplitude, and voxel signal phase; and calculating the wireless channel response at the prediction point coordinates based on the voxel density characteristics, voxel signal amplitude, and voxel signal phase. However, the modeling accuracy of the proposed method is difficult to meet different requirements, resulting in low modeling efficiency. Summary of the Invention

[0006] In view of the defects in the prior art, the object of the present invention is to provide a method and system for fast wireless channel modeling based on voxelized radiation field.

[0007] According to the present invention, a fast wireless channel modeling method based on voxelized radiation field is provided, comprising:

[0008] Step S1: Divide the space to be modeled into a three-dimensional voxel grid, and store the volume density and eigenvector of each voxel; sample the rays along the direction from the receiver Rx to the target to obtain multiple sampling points, and obtain the volume density and eigenvector of each sampling point through trilinear interpolation;

[0009] Step S2: Input the eigenvectors of the transmitter Tx position and the sampling point into the deformation network, and output the corrected eigenvalues. The radiation network combines the corrected eigenvalues ​​with the ray direction information to predict the signal strength transmitted from the sampling point to Rx.

[0010] Step S3: Calculate the cumulative received signals of all sampling points on each ray Rx based on the volume rendering model. The cumulative received signals are obtained by weighted summation, where the weights are determined by the transmittance of the sampling points and the signal termination probability.

[0011] Step S4: Taking the error between the spatial spectrum prediction value and the true observation value and the weighted sum of the background entropy loss function as the optimization target, the model is trained to approximate the true spatial spectrum.

[0012] Preferably, in step S1, the resolution of the three-dimensional voxel grid is gradually increased during the training process: a low-resolution grid is initially used, and is gradually upsampled to the target resolution by interpolation.

[0013] Preferably, in step S1, a spatial skipping mechanism is introduced into the sampling process: if the density value of a sampling point is lower than a set threshold, the sampling point is skipped in training and inference.

[0014] Preferably, in step S2, position coding is used to perform high-frequency feature expansion on the transmitter Tx position, sampling point position, and viewing angle direction. The position coding formula is:

[0015] E(p)=(sin(2 0 πp),cos(2 0 πp),…,sin(2 L-1 πp),cos(2 L-1 πp))#

[0016] where E(·) operates on the input vector element by element, each component is encoded independently, and L is a hyperparameter that controls the encoding dimension.

[0017] Preferably, in step S2, the deformable network and the radiation network are lightweight multi-layer perceptrons, wherein the deformable network output is used to correct the feature vector to decouple the influence of the static characteristics of the scene and the dynamic position of the Tx, and the radiation network outputs the direction-related signal strength.

[0018] Preferably, in step S3, the calculation method of the accumulated received signal is:

[0019]

[0020] α i =1-exp(-σ(x i )δ i )#

[0021]

[0022] Where K is the number of sampling points along the ray, w i =T i α i is the weight of the i-th virtual Tx, α i and α j are the probabilities that the signal terminates at the i-th and j-th sampling points, T i From x i Cumulative transmittance to Rx, δ i is the distance between adjacent sampling points.

[0023] Preferably, in step S4, the background entropy loss function is:

[0024]

[0025] in is a small batch of ray sets, T K (ω) represents the cumulative transmittance along ray ω;

[0026] The total loss function is:

[0027] L=L spectrum +λ bg L bg

[0028] Among them L spectrum is the mean square error of the spatial spectrum, λ bg is the weight coefficient;

[0029]

[0030] Where |Ω|=M×N represents the angular resolution of the spatial spectrum, corresponding to M azimuth angles φ and N elevation angles θ uniformly sampled on the unit upper hemisphere of Rx.

[0031] Preferably, in step S1, the local attributes encoded by the feature vector include: electrical conductivity and dielectric constant.

[0032] According to the present invention, a fast wireless channel modeling system based on voxelized radiation field is provided, comprising:

[0033] Module M1: Divides the space to be modeled into a 3D voxel grid, storing the volume density and eigenvector for each voxel; samples the ray along the direction from the receiver Rx to the target to obtain multiple sampling points, and obtains the volume density and eigenvector of each sampling point through trilinear interpolation;

[0034] Module M2: Inputs the eigenvectors of the transmitter Tx position and the sampling point into the deformation network and outputs the corrected eigenvalues. The radiation network combines the corrected eigenvalues ​​with the ray direction information to predict the signal strength transmitted from the sampling point to the Rx.

[0035] Module M3: Calculates the cumulative received signals of all sampling points on each ray Rx based on the volume rendering model. The cumulative received signals are obtained by weighted summation, where the weights are determined by the transmittance of the sampling points and the signal termination probability.

[0036] Module M4: The model is trained to approximate the true spatial spectrum by taking the error between the predicted spatial spectrum and the true observed value and the weighted sum of the background entropy loss function as the optimization target.

[0037] Preferably, in the module M1, the resolution of the three-dimensional voxel grid is gradually improved during the training process: a low-resolution grid is initially used, and is gradually upsampled to the target resolution by interpolation.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. By combining spatial voxel modeling with a dual-network decoupling structure, this method can efficiently fit the spatial spectral distribution of channels at any transmitter location, with excellent performance in complex multipath environments. Compared to traditional models based on ray tracing or unified function approximation, this method offers significant advantages in modeling accuracy, generalization, and training speed, and has great potential for engineering implementation.

[0040] 2. This paper provides a fast and efficient wireless channel modeling method that can accurately predict the received spatial spectrum and RSSI when the transmitter is moving. By voxelizing the radiation field and decoupling static environmental characteristics from dynamic Tx effects with the help of a deformable network, it achieves state-of-the-art performance on both RFID and BLE datasets. At the same time, the proposed method has the shortest training time and near-real-time inference speed.

[0041] 3. The present invention further improves the convergence speed and robustness of the model under sparse measurement conditions by introducing progressive learning, spatial skipping strategy and background entropy loss term. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0043] Figure 1 Schematic diagram of the reflection, diffraction and scattering of electromagnetic waves by environmental objects during propagation in the present invention;

[0044] Figure 2 Schematic diagram of virtual transmission point sampling along the receiving direction in the present invention;

[0045] Figure 3 Schematic diagram of obtaining volume density and features by trilinear interpolation of sampling points in the present invention;

[0046] Figure 4 This is a structural block diagram of a deformable network and a radial network according to an embodiment of the present invention;

[0047] Figure 5 The method and NeRF of the embodiment of the present invention are shown in the present invention. 2 Visual comparison of spatial spectrum reconstruction effects with WRF-GS+;

[0048] Figure 6 The present invention shows a comparison diagram of the cumulative distribution function (CDF) of the method described in the embodiment of the present invention and the existing method on the structural similarity (SSIM);

[0049] Figure 7The performance comparison diagram of the method described in the embodiment of the present invention and the existing method under different training set sizes is shown in the present invention;

[0050] Figure 8 The method and NeRF of the embodiment of the present invention are shown in the present invention. 2 ,Comparison chart of RSSI prediction error of MRI on BLE dataset. DETAILED DESCRIPTION

[0051] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0052] Definitions of Acronyms and Key Terms:

[0053] 6G Sixth Generation Mobile Communications Sixth Generation Mobile Communications System

[0054] NeRF Neural Radiance Field

[0055] Rx Receiver

[0056] Tx Transmitter

[0057] MSE Mean Square Error

[0058] WRF Wireless Radiance Field

[0059] GS Gaussian Splatting Gaussian Sputtering

[0060] SSIM Structural Similarity Index Structural Similarity Index

[0061] CDF Cumulative Distribution Function

[0062] RSSI Received Signal Strength Indicator Received Signal Strength Indicator

[0063] BLE Bluetooth Low Energy

[0064] MRI Model-driven Radio Interpolation

[0065] The present invention provides a fast wireless channel modeling method based on voxelized radiation fields, VoxelR. This method is applicable to scenarios with fixed receivers (Rx) and multi-position transmitters (Tx). It achieves efficient spatial spectrum synthesis through spatial directional sampling and signal accumulation. The method includes the following steps:

[0066] In the first step, for each transmitter, the system obtains the corresponding spatial angular power spectrum through ray sampling. Rays from the receiver to the specified direction are sampled uniformly to form multiple sampling points. Each sampling point is considered a virtual transmitter, and its spatial position is determined by the receiver position, ray direction, and sampling distance.

[0067] The second step is to calculate the volume density and radiation signal strength at each sampling point. Based on the volume rendering model, the transmittance and attenuation factor of the ray sampling point are calculated, and the signal contribution of each sampling point to the receiver is accumulated. Signal accumulation uses a weighted summation method, with the weights determined by the transmittance and termination probability.

[0068] The third step is to combine the accumulated signals along all ray directions to obtain the overall received signal at the receiver. The spatial spectrum distribution of this signal is calculated to model the spatial characteristics of the wireless channel.

[0069] In the fourth step, the error between the predicted value and the true observed value of the spatial spectrum is used as the optimization target, and the model is trained to approximate the true spatial spectrum.

[0070] The present invention proposes an explicit-implicit hybrid wireless signal radiation field representation method based on voxel grid, which effectively improves the training and inference speed of channel modeling.

[0071] In a specific embodiment, in the first step, the space to be modeled is divided into a three-dimensional voxel grid of fixed resolution, and a density value is stored in each voxel to represent the possibility of the location becoming a valid signal source and a feature vector to encode the local material or propagation environment properties. For each sampling point on a propagation ray, its density and characteristics are obtained by trilinear interpolation. The model initially uses a voxel grid with a coarser resolution to represent the signal field, and then gradually improves the grid resolution during the training process, upsampling the original voxels each time by interpolation until the target accuracy is achieved. This strategy can reduce the risk of overfitting and speed up convergence. In addition, a spatial skipping mechanism is introduced in the sampling process. If the density value of a sampling point is lower than the set threshold, the point will be skipped in subsequent training and inference to reduce redundant calculations during training and inference and save resources.

[0072] In the second step, to improve the model's ability to account for variations in transmitter Tx position, two lightweight MLPs are introduced: a deformable network and a radiating network. The deformable network takes the Tx position and feature vector as input and outputs a modified feature value. This effectively decouples the static characteristics of the scene from the radiation field variations caused by the dynamic Tx position, improving generalization of Tx positions. The radiating network combines the modified features with directional information to predict the signal strength transmitted from that point to the receiver. Furthermore, position encoding technology is used to perform high-frequency feature expansion on the transmitter position, sampling point location, and ray direction to capture the rapid spatial variations of wireless signals.

[0073] In the fourth step, a background entropy loss function is introduced to help the model clearly judge ray occlusion situations and accelerate convergence. The final training goal is to minimize the weighted sum of the combined spatial spectrum error and background entropy loss.

[0074] The embodiments of the present invention are described in further detail below.

[0075] The channel modeling method described in the present invention is suitable for scenarios with high requirements on deployment speed and real-time performance. Figure 1 As shown in , electromagnetic waves will undergo various interactions such as reflection, diffraction, and scattering on the surface of objects during propagation, making these surfaces become new signal radiation sources. Therefore, the receiver can receive signals from multiple directions. In order to calculate the signal received by Rx from a specific direction ω0, we sample along this direction and regard each sampling point as a virtual transmitter, as shown in Figure 2 As shown. We define the position x of the i-th virtual Tx i for

[0076] x i =P Rx +r i ω0#(1)

[0077] Among them, P Rx Indicates the location of the receiving end, r i is the distance from Rx to the i-th sampling point.

[0078] We use two voxel grids to represent the radiation field: the density voxel grid and feature voxel grid Among them (L x ,L y ,L z ) is the size of the voxel grid, and F is the feature space dimension. For each sample X i ,like Figure 3 As shown, we obtain its volume density σ(X i ) and the eigenvector Feat(X i ):

[0079]

[0080] The volume density σ(x i ) represents the probability that the point is a virtual Tx due to the propagation effect. i ) depends only on spatial position and encodes physical parameters such as conductivity and dielectric constant.

[0081] like Figure 4 As shown in Figure 2, we use two shallow MLPs to predict the direction-dependent signal radiated by each sampling point. The deformation network D is used to model the influence of Tx position on local multipath characteristics and output the deformation feature ΔFeat(x i ), thereby realizing the modeling of the radiation field related to the Tx position. Subsequently, the deformation features and the original features are fused, and the radiation signal S(x i ,-ω0). This method decouples the dynamic and static radiation fields caused by the movement of Tx position by introducing the deformation network D, thereby improving the generalization ability of the model. In addition, we T , sampling point position x i and viewing direction -ω0 apply position encoding to capture high-frequency variations:

[0082] E(p)=(sin(2 0 πp),cos(2 0 πp),…,sin(2 L-1 πp),cos(2 L-1 πp))#(4)

[0083] where E(·) operates on the input vector element by element, each component is encoded independently, and L is a hyperparameter that controls the encoding dimension.

[0084] For each direction ω0, the cumulative received signal R(ω0) is calculated using the following volume rendering formula:

[0085]

[0086] α i =1-exp(-σ(x i )δ i )#(6)

[0087]

[0088] Where K is the number of sampling points along the ray, w i =T i α i is the weight of the i-th virtual Tx, α i and α jare the probabilities that the signal terminates at the i-th and j-th sampling points, T i From x i Cumulative transmittance to Rx, δ i is the distance between adjacent sampling points. For the spatial spectrum prediction task, the training objective is to minimize the mean square error (MSE) of the signal prediction error in each direction:

[0089]

[0090] Where |Ω|=M×N represents the angular resolution of the spatial spectrum, corresponding to M azimuth angles φ and N elevation angles θ uniformly sampled on the unit upper hemisphere of Rx.

[0091] We adopt the following three training strategies to optimize performance and accelerate training convergence and inference speed.

[0092] First, a progressive learning strategy is adopted. In the initial stage of training, a higher resolution voxel grid with the number of voxels is used. The grid resolution is gradually increased in M ​​stages. In each stage, the number of voxels is doubled and V is interpolated by trilinear interpolation. density and V feature Upsampling is performed until the final resolution (L x ,L y ,L z ). This strategy can reduce the risk of the algorithm falling into local optimality and improve the stability of training.

[0093] Furthermore, in real wireless propagation environments, most of the space is air, which contributes minimally to the received signal. Using a uniform sampling strategy would result in a significant amount of inefficient computation. Therefore, to improve sampling efficiency, we introduce a spatial skipping strategy during training and inference. If the volume density of a sampling point falls below a set threshold τ, indicating that it is in air, that point is skipped during subsequent training and inference.

[0094] Finally, we introduce the background entropy loss function L bb , used to improve the ability to identify occlusion situations:

[0095]

[0096] in is a small batch of ray sets, T K (ω) represents the cumulative transmittance along ray ω. This loss encourages the network to make clearer predictions about the occluded area, thereby accelerating training convergence. The final total loss function is

[0097] L=L spectrum +λ bg L bg

[0098] Among them L spectrum From formula (8), λ bg is the weight coefficient.

[0099] To verify the performance of the voxelized radiation field proposed in this paper in real-world scenarios, we conducted simulation experiments using a publicly available wireless channel dataset. The experiments examined various aspects, including spatial spectrum reconstruction performance, image structure similarity, data scale robustness, training and inference efficiency, and received signal strength indicator (RSSI) prediction. The details are as follows:

[0100] The experiment used NeRF 2 This dataset is an open-source RFID dataset. This dataset was collected in an indoor experimental environment measuring 23m × 24m × 2m. The receiver is fixed and equipped with a 4×4 antenna array operating at 915MHz. The transmitter is a movable RFID tag, randomly placed in space, continuously transmitting an RN16 signal. The dataset contains 6,123 transmitter locations, each corresponding to a 360×90 spatial spectrum representing the received signal strength distribution at that location in all directions. Unless otherwise specified, 80% of the data is used for training and 20% for testing.

[0101] In order to verify the effectiveness of the proposed method, it is compared with the following two existing methods: NeRF, which is the first method to use neural radiation field for wireless channel modeling. 2 and the wireless radiation field (WRF) modeling method WRF-GS+ based on 3D Gaussian sputtering (GS).

[0102] Figure 5 The method of the present invention is combined with NeRF 2 The spatial spectrum synthesized by WRF-GS+ at multiple transmitting end positions is visually compared with the real spectrum. It can be clearly seen that the spatial spectrum generated by the method proposed in the present invention is closest to the real image, and the transition between high-intensity and low-intensity areas is clearer, indicating that it can better retain high-frequency information. This shows that the method of the present invention can more accurately capture the rapid spatial fluctuations caused by small-scale fading effects. In contrast, although WRF-GS+ is better than NeRF 2 It has better performance, but because it uses an essentially smooth and continuous Gaussian, the edges of the generated image are blurred, especially in the boundary area, and the sharp details of the image are lost.

[0103] Figure 6The structural similarity index (SSIM) was further used to evaluate the similarity between the synthetic image and the real spatial spectrogram. The cumulative distribution function (CDF) results showed that the median SSIM of the proposed method was 0.9035, which was better than WRF-GS+ (0.8813) and NeRF 2 (0.7991). This paper introduces a deformation network to separate dynamic and static fields, enhancing the ability to model changes in mobile transmitters and significantly improving spatial spectrum synthesis performance. In contrast, the original NeRF model is used to model static radiation fields, and therefore has limited ability to perceive radiation field changes caused by changes in transmitter position.

[0104] Figure 7 The training data was set to 20%, 40%, 60%, 80%, and 100% respectively, and the performance of each method under different training set sizes was compared. The results show that under all data sizes, the proposed scheme outperforms the comparison method and shows good robustness to sparse data. At the same time, the model has tended to converge when using about 60% of the training data, reflecting its excellent data efficiency. This is because compared with NeRF 2 Compared with the purely implicit representation in

[15] , the present invention adopts an explicit voxel grid to represent the radiation field and ensures local continuity through interpolation, which plays a regularization role when the data is sparse and reduces the risk of overfitting.

[0105] We further compared the training and inference time of each method. 2 The training time is 15 hours and 30 minutes, while WRF-GS+ takes 2 hours and 40 minutes. The algorithm proposed in this paper takes only 20 minutes, which is 8 times faster than WRF-GS+ and 10 times faster than NeRF. 2 Nearly 47 times faster. In the inference stage, the inference time of the present invention is 0.042 seconds, which is about 2 1 / 10 of that (0.396 seconds), which is close to WRF-GS+ (0.01 seconds). This is due to the fact that VoxelRF uses trilinear interpolation of voxel grid instead of NeRF. 2 The deep MLP network in

[15] is used and the sampling points with low volume density are skipped. The above results show that the proposed method is very suitable for deployment in scenarios that are sensitive to latency or have limited resources.

[0106] Considering that the spatial spectrum can be obtained by integrating the directional power to obtain the RSSI value, further NeRF 2 The performance of this method in the RSSI prediction task is tested on the open-source BLE dataset in [1]. In the BLE dataset, 21 receivers and one mobile transmitter are deployed in a 15,000 square foot indoor scene, containing 2,000 valid RSSI measurements (greater than -100dBm). The results are shown in the figure. Figure 8As shown in Figure 2, the median error of RSSI prediction of the proposed solution is 2.66dB, which is better than NeRF 2 (2.85dB) and the model-driven wireless interpolation method MRI (6.74dB). Overall, the combined algorithm proposed in the present invention has better performance.

[0107] The present invention also provides a fast wireless channel modeling system based on voxelized radiation field. The fast wireless channel modeling system based on voxelized radiation field can be implemented by executing the process steps of the fast wireless channel modeling method based on voxelized radiation field. That is, those skilled in the art can understand the fast wireless channel modeling method based on voxelized radiation field as a preferred implementation of the fast wireless channel modeling system based on voxelized radiation field.

[0108] A fast wireless channel modeling system based on voxelized radiation field, comprising:

[0109] Module M1: Divides the space to be modeled into a 3D voxel grid, storing the volume density and eigenvector for each voxel; samples the ray along the direction from the receiver Rx to the target to obtain multiple sampling points, and obtains the volume density and eigenvector of each sampling point through trilinear interpolation;

[0110] Module M2: Inputs the eigenvectors of the transmitter Tx position and the sampling point into the deformation network and outputs the corrected eigenvalues. The radiation network combines the corrected eigenvalues ​​with the ray direction information to predict the signal strength transmitted from the sampling point to the Rx.

[0111] Module M3: Calculates the cumulative received signals of all sampling points on each ray Rx based on the volume rendering model. The cumulative received signals are obtained by weighted summation, where the weights are determined by the transmittance of the sampling points and the signal termination probability.

[0112] Module M4: The model is trained to approximate the true spatial spectrum by taking the error between the predicted spatial spectrum and the true observed value and the weighted sum of the background entropy loss function as the optimization target.

[0113] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0114] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A fast wireless channel modeling method based on voxelized radiation field, characterized in that: include: Step S1: Divide the space to be modeled into a three-dimensional voxel grid, and store the volume density and feature vector for each voxel; Multiple sampling points are obtained by sampling along the ray from the receiver Rx to the target, and the volume density and eigenvector of each sampling point are obtained by trilinear interpolation; Step S2: Input the eigenvectors of the transmitter Tx position and the sampling point into the deformation network, and output the corrected eigenvalues. The radiation network combines the corrected eigenvalues ​​with the ray direction information to predict the signal strength transmitted from the sampling point to Rx. Step S3: Calculate the cumulative received signals of all sampling points on each ray Rx based on the volume rendering model. The cumulative received signals are obtained by weighted summation, where the weights are determined by the transmittance of the sampling points and the signal termination probability. Step S4: Taking the error between the spatial spectrum prediction value and the true observation value and the weighted sum of the background entropy loss function as the optimization target, the model is trained to approximate the true spatial spectrum.

2. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S1, the resolution of the three-dimensional voxel grid is gradually increased during the training process: a low-resolution grid is initially used, and the grid is gradually upsampled to the target resolution through interpolation.

3. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S1, a spatial skipping mechanism is introduced into the sampling process: if the density value of a sampling point is lower than a set threshold, the sampling point is skipped during training and inference.

4. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S2, position coding is used to perform high-frequency feature expansion on the transmitter Tx position, sampling point position, and viewing angle direction. The position coding formula is: E(p)=(sin(2 0 πp),cos(2 0 πp),…,sin(2 L-1 πp),cos(2 L-1 πp))# where E(·) operates on the input vector element by element, each component is encoded independently, and L is a hyperparameter that controls the encoding dimension.

5. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S2, the deformable network and the radiation network are lightweight multi-layer perceptrons, wherein the deformable network output is used to correct the feature vector to decouple the influence of the static characteristics of the scene and the dynamic position of the Tx, and the radiation network outputs the direction-related signal strength.

6. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S3, the calculation method of the accumulated received signal is: a i =1-exp(-σ(x i )d i )# Where K is the number of sampling points along the ray, w i =T i α i is the weight of the i-th virtual Tx, α i and α j are the probabilities that the signal terminates at the i-th and j-th sampling points, T i From x i Cumulative transmittance to Rx, δ i is the distance between adjacent sampling points.

7. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In step S4, the background entropy loss function is: in is a small batch of ray sets, T K (ω) represents the cumulative transmittance along ray ω; The total loss function is: L=L spectrum +λ bg L bg Among them L spectrum is the mean square error of the spatial spectrum, λ bg is the weight coefficient; Where |Ω|=M×N represents the angular resolution of the spatial spectrum, corresponding to M azimuth angles φ and N elevation angles θ uniformly sampled on the unit upper hemisphere of Rx.

8. The method for rapid wireless channel modeling based on voxelized radiation field according to claim 1, characterized in that: In the step S1, the local attributes encoded by the feature vector include: conductivity and dielectric constant.

9. A fast wireless channel modeling system based on voxelized radiation field, characterized in that: include: Module M1: Divide the space to be modeled into a 3D voxel grid, and store the volume density and eigenvector of each voxel; Multiple sampling points are obtained by sampling along the ray from the receiver Rx to the target, and the volume density and eigenvector of each sampling point are obtained by trilinear interpolation; Module M2: Inputs the eigenvectors of the transmitter Tx position and the sampling point into the deformation network and outputs the corrected eigenvalues. The radiation network combines the corrected eigenvalues ​​with the ray direction information to predict the signal strength transmitted from the sampling point to the Rx. Module M3: Calculates the cumulative received signals of all sampling points on each ray Rx based on the volume rendering model. The cumulative received signals are obtained by weighted summation, where the weights are determined by the transmittance of the sampling points and the signal termination probability. Module M4: The model is trained to approximate the true spatial spectrum by taking the error between the predicted spatial spectrum and the true observed value and the weighted sum of the background entropy loss function as the optimization target.

10. The fast wireless channel modeling system based on voxelized radiation field according to claim 9, characterized in that: In the module M1, the resolution of the three-dimensional voxel grid is gradually improved during the training process: a low-resolution grid is initially used and is gradually upsampled to the target resolution through interpolation.

Images