A method for 3D environment feature extraction and electromagnetic propagation path modeling based on neural networks

Abstract

This invention discloses a method for 3D environmental feature extraction and electromagnetic propagation path modeling based on neural networks, belonging to the field of channel modeling technology. The method includes: S1: acquiring wireless channel data of the target scene and corresponding 3D spatial point cloud environmental data; S2: obtaining a physical environment feature dataset for neural network training; S3: constructing a 3D environmental feature extraction model to generate implicit environmental feature representations reflecting the relationship between environmental spatial distribution and material properties; S4: constructing an electromagnetic propagation direction prediction model and a power prediction model to predict the step-by-step evolution of the electromagnetic propagation path and output electromagnetic propagation path-level characteristics. This invention solves the problems of complex and inefficient existing modeling processes, difficulty in directly utilizing measured point cloud data, and difficulty in efficiently generating multipath propagation paths. This invention improves modeling efficiency and reduces labor costs, enabling direct modeling of point cloud geometry and material properties while reducing the computational complexity of multipath path inference.

Description

Technical Field

[0001] This invention relates to the field of channel modeling technology, specifically to a method for three-dimensional environmental feature extraction and electromagnetic propagation path modeling based on neural networks. Background Technology

[0002] With the continuous evolution of 5G and future 6G networks, complex communication scenarios pose unprecedented challenges to the accuracy and capability of channel modeling. Accurately simulating the propagation path of wireless signals in complex, dynamically changing environments has become a key challenge in modern communication system design. The propagation process of electromagnetic waves is constrained by various factors, including terrain, buildings, materials, and dynamic environmental changes. These factors not only affect the signal propagation range but also lead to complex effects such as multipath propagation, signal attenuation, reflection, and refraction, greatly increasing the difficulty of channel modeling. Against this backdrop, deterministic channel modeling methods that reconstruct propagation paths based on real-world environments have attracted increasing attention. By physically characterizing the propagation mechanism, they possess significantly higher interpretability and predictive capabilities than statistical models and have gradually become a research hotspot in the field of wireless propagation.

[0003] Traditional deterministic channel modeling methods often rely on empirical models or physical models based on simplified assumptions. While these methods can meet the needs of specific environments to a certain extent, they often fail to accurately reflect the true electromagnetic propagation characteristics in complex environments (such as cities, railways, and industrial sites), resulting in bottlenecks in both the accuracy and efficiency of signal modeling.

[0004] To address the aforementioned issues, neural network-based channel modeling methods have gradually gained widespread attention. Neural networks have significant advantages in processing large-scale data and high-dimensional features. They can automatically learn key information such as geometric structures, surface morphology, and material properties in a three-dimensional environment through data-driven methods, and combine this with electromagnetic propagation mechanisms to model wireless channels, thereby improving the accuracy and efficiency of channel simulation to a certain extent.

[0005] Most current mainstream neural network-based modeling techniques rely on mesh structures as the representation of the environment, typically requiring the pre-construction of a complete 3D physical model. These methods still face challenges in complex or dynamic scenes, including cumbersome modeling processes, high computational costs, and difficulties in scene updates, limiting their flexibility and efficiency in engineering applications. Specific drawbacks are as follows:

[0006] 1) Existing deterministic channel modeling methods rely heavily on manually constructed 3D physical models. The modeling process is complex and inefficient. Existing deterministic modeling methods such as ray tracing usually require the construction of environmental geometry based on CAD models or meshed models and the manual configuration of material parameters. In complex or irregular environments, the construction process of such models is time-consuming and labor-intensive, and it is difficult to adapt to changes in environmental structure quickly, which limits their application efficiency in practical engineering.

[0007] 2) Existing neural network-based propagation modeling methods mostly rely on gridded environment representations, making it difficult to directly utilize measured point cloud data. Currently, mainstream neural network modeling techniques typically use regular grids or explicit 3D models as inputs, requiring point cloud data to be reconstructed and gridded first, which increases the complexity of data preprocessing. At the same time, when point clouds contain noise, uneven density, or local missing data, the grid reconstruction process is prone to introducing errors, affecting the accuracy of subsequent propagation modeling.

[0008] 3) Existing deterministic or semi-deterministic modeling methods have high computational complexity in complex scenarios, making it difficult to efficiently generate multipath propagation paths. Traditional ray tracing methods require enumerating and searching a large number of candidate rays. As the propagation order increases, the number of paths grows exponentially, leading to a significant increase in computational overhead. Even with the introduction of some learning methods, they still generally rely on explicit path search or intermediate geometric judgments, making it difficult to achieve efficient multipath modeling while ensuring accuracy. Summary of the Invention

[0009] The purpose of this invention is to provide a method for three-dimensional environmental feature extraction and electromagnetic propagation path modeling based on neural networks. This method can extract propagation environment features based on point cloud data, which is beneficial for efficient modeling of electromagnetic propagation paths. This provides a reference for channel modeling and optimization of complex wireless communication systems and solves the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for 3D environment feature extraction and electromagnetic propagation path modeling based on neural networks, comprising:

[0012] S1: Deploy a wireless channel measurement system and a synchronized 3D point cloud environment acquisition system to acquire wireless channel data and corresponding 3D spatial environment point cloud data of the target scene;

[0013] S2: Preprocess the wireless channel data and point cloud data to obtain a physical environment feature dataset for neural network training;

[0014] S3: Construct a three-dimensional environmental feature extraction model based on the physical environment feature dataset to generate an implicit environmental feature representation that reflects the relationship between environmental spatial distribution and material properties;

[0015] S4: Based on the implicit environmental features, construct an electromagnetic propagation direction prediction model and a power prediction model, use a neural network to output multipath parameters hop by hop, predict the step-by-step evolution process of the electromagnetic propagation path, and output the electromagnetic propagation path-level characteristics.

[0016] Preferably, in step S1, the deployment of a wireless channel measurement system and a synchronized three-dimensional point cloud environment acquisition system includes:

[0017] The target environment is divided into Region 1 and Region 2 based on differences in spatial structure.

[0018] Among them, Region 1 and Region 2 have similar environmental characteristics but different physical structure layouts. The data in Region 1 is divided into training set and test set according to a preset ratio, and all the data in Region 2 is used as test set to evaluate the generalization ability of the model.

[0019] During the actual test, the lidar and the signal receiver were deployed synchronously on the same mobile platform, while the signal transmitter was placed at a fixed location.

[0020] At the start of the measurement, the signal receiver continuously receives and stores the vector detection signal, while the lidar synchronously collects and records the three-dimensional structural data of the environment in real time, realizing the time-synchronous storage of wireless channel data and three-dimensional spatial environment point cloud data.

[0021] Preferably, in step S2, the wireless channel data and point cloud data are preprocessed to obtain a physical environment feature dataset for neural network training, including:

[0022] The wireless channel data is noise-removed, and the channel transfer function and channel impulse response are obtained by performing a fast Fourier transform on the channel sounding signal. Multipath propagation parameters, including multipath delay, angle of arrival, received power and path loss, are extracted.

[0023] The point cloud data is subjected to coordinate alignment, noise reduction filtering and downsampling processing, and the point cloud normal vector is estimated based on neighborhood features to construct an environmental feature display representation that includes geometric structure, surface morphology and material properties.

[0024] Preferably, in step S3, a three-dimensional environmental feature extraction model is constructed based on the physical environment feature dataset to generate an implicit environmental feature representation reflecting the relationship between environmental spatial distribution and material properties, including:

[0025] Fourier position encoding is performed on the point cloud coordinates to map the three-dimensional geometric information and material properties to a high-dimensional feature space;

[0026] The PointNet++ network is used to extract multi-level features from the encoded point cloud. The PointNet++ network obtains local geometric features that can characterize the spatial structure at different scales by performing layer-by-layer sampling, neighborhood search and local feature aggregation on the point cloud.

[0027] By constructing an implicit environmental feature expression that can simultaneously reflect the spatial structure, surface properties, and material properties of the environment through global feature fusion and nonlinear mapping structure, a well-trained electromagnetic environment feature extraction model can be obtained for subsequent electromagnetic propagation path modeling.

[0028] Preferably, in step S4, constructing an electromagnetic propagation direction prediction model based on the implicit environmental characteristics includes:

[0029] An electromagnetic propagation direction prediction model based on a Transformer network is constructed. The Transformer encoder adopts a multi-head self-attention mechanism and a feedforward neural network, which includes 4 attention heads and 3 encoding layers, and outputs a three-dimensional direction vector.

[0030] The Transformer network is trained using a dataset, with implicit environmental feature data as input;

[0031] The loss function is constructed as the direction vector similarity between the true value and the predicted value, which is used to optimize the model performance. The direction vector similarity is measured by calculating the cosine similarity between the predicted direction vector and the actual direction vector.

[0032] The Transformer network is trained using this loss function to obtain a trained Transformer model, which is used as an electromagnetic propagation direction prediction network to infer the next hop propagation direction.

[0033] Preferably, in step S4, constructing a power prediction model based on the implicit environmental features includes:

[0034] The power prediction model is based on a multilayer perceptron network. The network contains four fully connected layers with layer sizes of 128, 256, 64, and 8, respectively.

[0035] The network uses the ReLU activation function for nonlinear transformation, resulting in an 8-dimensional vector as the decay matrix for each hop.

[0036] Preferably, in step S4, the stepwise evolution process of the electromagnetic propagation path is predicted by setting a maximum number of bounces to limit the iterative process of the propagation path, thereby obtaining an electromagnetic propagation path prediction model.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] This invention extracts environmental features directly from measured point cloud environmental data, avoiding the reliance on manual modeling or complex mesh reconstruction in traditional modeling processes. This simplifies the modeling process, improves modeling efficiency in complex environments, and reduces labor costs. It uses neural networks to extract and encode features from point cloud environmental data, integrating spatial structure, surface morphology, and material properties to construct an implicit environmental feature representation. This provides a more complete and accurate environmental input for electromagnetic propagation path modeling, enabling direct modeling of point cloud geometry and material properties. By predicting electromagnetic propagation direction and power attenuation step-by-step through neural networks, it avoids exhaustive search of the entire path, significantly reducing computational complexity while maintaining propagation modeling accuracy. This achieves efficient modeling of electromagnetic propagation paths in complex scenarios and reduces the computational complexity of multipath inference. Attached Figure Description

[0039] Figure 1 This is a flowchart of the three-dimensional environment feature extraction and electromagnetic propagation path modeling method of the present invention;

[0040] Figure 2 This is a schematic diagram of the structure for detecting environmental data in the lobbies of different teaching buildings according to the present invention.

[0041] Figure 3 A flowchart illustrating the point cloud processing of three rooms in different teaching building lobbies according to the present invention;

[0042] Figure 4 This is a diagram of the PointNet++ network structure for multi-level feature extraction of point clouds according to the present invention.

[0043] Figure 5 This is a diagram of the Transformer network structure for constructing the electromagnetic propagation direction prediction model of the present invention.

[0044] Figure 6 This is a flowchart of the network architecture for predicting the step-by-step evolution of electromagnetic propagation paths according to the present invention.

[0045] Figure 7 This is a comparison diagram of the predicted propagation results of electromagnetic waves of different orders in Hall 1 of the present invention and the actual channel impulse response.

[0046] Figure 8 This is a diagram showing the generalization result of the trained model of this invention in the Hall 2 scene.

[0047] Figure 9 This is a diagram showing the generalization result of the trained model of this invention in the Hall 3 scene.

[0048] Figure 10 This is a comparison diagram of the predicted results and true values ​​of electromagnetic wave propagation of different orders according to the present invention.

[0049] Figure 11 This is a comparison diagram of the prediction results of different orders and the true values ​​under the power change trend of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To address the issues of complex and inefficient existing modeling processes, difficulty in directly utilizing measured point cloud data, and challenges in efficiently generating multipath propagation paths, please refer to [link to relevant documentation]. Figures 1-11 This embodiment provides the following technical solution:

[0052] A method for 3D environment feature extraction and electromagnetic propagation path modeling based on neural networks, comprising:

[0053] S1: Deploy a wireless channel measurement system and a synchronized 3D point cloud environment acquisition system to acquire wireless channel data and corresponding 3D spatial environment point cloud data of the target scene.

[0054] It should be noted that point cloud data refers to a set of data that represents the spatial structure of the environment in the form of discrete points, obtained through three-dimensional sensing devices such as LiDAR. Each point contains at least three-dimensional coordinate information and can be further associated with additional features such as normal vectors and material properties.

[0055] In this embodiment, a wireless channel measurement system and a three-dimensional point cloud environment acquisition system are set up in the lobby of the teaching building to obtain wireless channel data of the target scene and corresponding three-dimensional spatial environment point cloud data. By selecting specific points in the lobby of the teaching building for synchronous detection, the channel propagation characteristics and environmental spatial features are collected in real time to ensure the accurate correspondence between channel data and environmental point cloud data.

[0056] In this embodiment, the deployment of a wireless channel measurement system and a synchronized three-dimensional point cloud environment acquisition system further includes: dividing the target environment into region one and region two according to differences in spatial structure; wherein, region one and region two have similar environmental characteristics but different physical structure layouts, the data of region one is divided into training set and test set according to a preset ratio, and all the data of region two is used as test set to evaluate the generalization ability of the model. It can be understood that, based on different environmental structures, the datasets of different regions are divided into training set, test set and validation set.

[0057] In this embodiment, during the actual measurement process, the lidar and the signal receiver are synchronously deployed on the same mobile platform, and the signal transmitter is placed at a fixed point. At the start of the measurement, the signal receiver continuously receives and stores the vector detection signal, and the lidar synchronously collects and records the three-dimensional structural data of the environment in real time, realizing the time-synchronous storage of wireless channel data and three-dimensional spatial environment point cloud data. It can be understood that synchronous measurement enables the physical environment and the electromagnetic environment to correspond in real time, so as to deeply analyze the relationship between electromagnetic wave propagation and environmental characteristics.

[0058] For example, environmental data was collected in the lobbies of different teaching buildings. Each lobby was 50m long, 17m wide, and 7m high, with variations in its internal structure. Figure 2 As shown, for Hall 1, 80% of the data was selected as the training set and 20% as the test set. Hall 2 kept its main structure unchanged, but the internal scattering wall and column structure changed. The internal structure of Hall 3 was significantly adjusted. All the data from Halls 2 and 3 were used as the validation set. This dataset partitioning method was used to test the model's generalization ability under different structural changes in similar scenarios.

[0059] S2: Preprocess the wireless channel data and point cloud data to obtain a physical environment feature dataset for neural network training.

[0060] In this embodiment, noise is removed from the wireless channel data. The channel transfer function and channel impulse response are obtained by performing a Fast Fourier Transform (FFT) on the channel sounding signal. Multipath propagation parameters such as multipath delay, angle of arrival, received power, and path loss are then extracted. For example, the acquired IQ signal is processed using FFT to obtain the channel impulse response. Redundant data in the snapshot is then removed, and the generalized spatial expectation-maximization algorithm is used to extract multipath information in the space. The redundant data refers to data from repeated points during the measurement process, as well as similar feature data caused by measurement errors or external interference.

[0061] In this implementation example, coordinate alignment, denoising filtering, and downsampling are performed on the point cloud data. Point cloud normal vectors are estimated based on neighborhood features to construct an environmental feature display representation that includes geometric structure, surface morphology, and material properties. For example, a point cloud filtering algorithm is used to remove noise points within the spatial range to ensure the accuracy of the point cloud data. Subsequently, the farthest point sampling method is used to reduce the number of spatial point clouds while preserving the details of the spatial structure, ensuring that the processed point cloud data can still effectively represent the geometric features of the environment. In this example, the number of point clouds is downsampled from 8,906,543 to 251,454 to reduce computational complexity. Based on the processed point cloud data, the surface normal vectors of the point cloud are calculated using the SVD decomposition method, and material feature attributes are assigned to individual points for subsequent feature input to the model. Then, based on the processed point cloud data, singular value decomposition is used to calculate the surface normal vectors of each point to accurately describe the surface features. These normal vectors, along with the corresponding point material features, serve as input features for subsequent neural network model training and modeling. The point cloud processing flow for the three rooms is as follows: Figure 3 As shown.

[0062] S3: Construct a three-dimensional environmental feature extraction model based on the physical environment feature dataset to generate an implicit environmental feature representation that reflects the relationship between environmental spatial distribution and material properties.

[0063] It should be noted that electromagnetic environment characteristics refer to environmental description information composed of the spatial geometry, surface morphology and material properties of the real physical environment, used to characterize the physical properties related to propagation path, reflection, scattering and other behaviors during electromagnetic wave propagation.

[0064] In this embodiment, Fourier position encoding is performed on the point cloud coordinates, and discrete three-dimensional geometric information and material properties are mapped to a high-dimensional feature space. It can be understood that the position encoding adopts a sinusoidal encoding method. By applying sine and cosine functions of different frequencies to the three-dimensional coordinates of the point cloud, the original coordinates are mapped from a low-dimensional space to a high-dimensional space with richer expressive power. Specifically, each coordinate component will undergo a series of sin / cos function transformations with increasing frequency to generate a set of features with multi-scale spatial sensitivity, enabling the model to perceive the geometric changes of the environment at different spatial scales.

[0065] In this embodiment, the PointNet++ network is used to perform multi-level feature extraction on the encoded point cloud. The PointNet++ network performs layer-by-layer sampling, neighborhood search, and local feature aggregation on the point cloud to obtain local geometric features that can characterize the spatial structure at different scales. It can be understood that the PointNet++ network, utilizing its positional invariance to point order and rotational invariance achieved through local neighborhood modeling, can robustly extract the spatial structure features of the point cloud without relying on a fixed grid or regular structure. For example, inputting the point cloud around the interaction point of the electromagnetic propagation path into the PointNet++ trained network yields its spatial structure feature encoding. The network structure is as follows: Figure 4 As shown.

[0066] In this embodiment, an implicit environmental feature representation that can simultaneously reflect the spatial structure, surface properties, and material properties of the environment is constructed through global feature fusion and a nonlinear mapping structure, resulting in a trained electromagnetic environment feature extraction model for subsequent electromagnetic propagation path modeling. It can be understood that the global feature fusion refers to concatenating and fusing the spatial structure encoding output by the PointNet++ network with high-dimensional material properties as a three-dimensional environmental feature representation.

[0067] For example, in the complex indoor environment shown in Hall 1, a transmitter radiates multiple electromagnetic wave propagation rays into space. During propagation, these rays are blocked and affected by scattering bodies such as walls and columns, resulting in various electromagnetic propagation modes, including reflection and scattering. To accurately describe the propagation characteristics of electromagnetic waves in a local environment, N points within the neighborhood of the interaction point between the ray and the obstacle are selected as a local spatial subset and input into the PointNet++ network to extract local geometric features. This network effectively characterizes the spatial morphology, surface orientation, and structural changes of the interaction area by performing hierarchical sampling and local feature aggregation on the neighborhood point cloud. Simultaneously, the material category information of the interaction points is mapped to a high-dimensional feature space using Fourier sinusoidal position encoding to enhance the network's ability to perceive differences in the electromagnetic responses of different materials. Subsequently, the encoded material features are concatenated with the local geometric features output by PointNet++ to form an implicit electromagnetic environment feature expression vector that includes local spatial structure, surface normal vector changes, and material properties.

[0068] S4: Based on the implicit environmental features, construct an electromagnetic propagation direction prediction model and a power prediction model, use a neural network to output multipath parameters hop by hop, predict the step-by-step evolution process of the electromagnetic propagation path, and output the electromagnetic propagation path-level characteristics.

[0069] It should be noted that multipath refers to the propagation path of electromagnetic waves in space.

[0070] In this embodiment, an electromagnetic propagation direction prediction model based on a Transformer network is constructed. The Transformer encoder employs a multi-head self-attention mechanism and a feedforward neural network, containing four attention heads and three encoding layers, outputting a three-dimensional direction vector. The Transformer network is trained using the dataset, with implicit environmental feature data as input. Furthermore, the loss function is constructed as the direction vector similarity between the ground truth and the predicted value to optimize model performance. This direction vector similarity is measured by calculating the cosine similarity between the predicted and actual direction vectors. It can be understood that the neural network input is an environmental feature vector, and the output is a single-order electromagnetic propagation direction. The network structure is as follows: Figure 5 As shown.

[0071] In this embodiment, a power prediction model based on a multilayer perceptron network is used. The network uses the ReLU activation function for nonlinear transformation to obtain an 8-dimensional vector as the attenuation matrix for each hop. It can be understood that the single-hop propagation attenuation of electromagnetic waves can be directly obtained through the network power prediction model, which is used to calculate the final power loss.

[0072] In this embodiment, the progressive evolution of the electromagnetic propagation path is predicted by setting a maximum number of bounces to control the iterative process of the propagation path, thus obtaining an electromagnetic propagation path prediction model. It is understood that in practical applications, the network needs to set the highest transmission order of the electromagnetic wave to determine propagation termination and obtain different electromagnetic propagation sequences.

[0073] For example, Figure 6 The entire network architecture process is demonstrated. In this example, firstly, during the ray emission stage, to obtain the initial incident direction with uniform sampling throughout space, Fibonacci spherical sampling is used to generate the direction vector. Compared to the traditional uniform grid sampling method, Fibonacci sampling has better consistency in spherical distribution density, avoiding the problem of excessive density in the polar regions or excessive sparseness in the equator. This experiment sets the number of emitted rays to 50,000, and each ray is allowed a maximum of 3 bounces during propagation to cover common propagation modes in the experimental environment, such as single reflection, wall diffraction, and multiple scattering. When a ray reaches the maximum number of bounces or encounters a strongly absorbing material, the path will be automatically terminated. The ray propagation modes are shown in Table 1.

[0074] Table 1: Configuration of Ray Propagation Modes

[0075]

[0076] The model uses 3D point cloud, material features, and the initial direction of ray propagation as network inputs. To enable the neural network to fully capture spatial location information and material differences in the 3D structure, a sinusoidal Fourier positional encoding is employed to map point cloud coordinates and material properties to a high-dimensional feature space. The model extracts a local sub-cloud of N=128 neighborhood points centered on the interaction point between the electromagnetic wave and the obstacle. Through global pooling and local aggregation, an implicit environmental feature encoding that simultaneously represents spatial structure, surface normal vector changes, and material distribution is obtained. In the propagation path inference stage, a hop-by-hop direction prediction network based on a Transformer architecture is constructed. This network takes the environmental feature encoding, the incident direction, and the position of the previous hop interaction point as inputs. It captures the influence of environmental geometry and material information on the propagation direction through a self-attention mechanism and outputs the predicted direction vector for the next hop. Simultaneously, the model utilizes a multilayer perceptron structure to predict the propagation attenuation matrix, used to represent power loss during reflection and scattering. During training, direction prediction employs a loss function based on "1-cosine similarity" to ensure that the angular error of the output direction is minimized; power prediction uses a dB-domain mean square error loss to improve the learning stability for power data with a large dynamic range. The experiment used the Adam optimizer with a learning rate of 1e-4, a batch size of 64, and a total training duration of 800 epochs. The specific network training parameters are shown in Table 2.

[0077] Table 2: Summary Table of Network Training Parameters

[0078]

[0079] After the complete hop-by-hop prediction is completed, the model generates a sequence of electromagnetic propagation paths that includes the location of the interaction point, reflection type, incident / outgoing direction, and path loss. For each ray, parameters such as total path loss, total delay, and angular spread are calculated based on the generated path sequence and aligned with the multipath output of the measured channel.

[0080] For example, Figure 7The comparison between the predicted propagation results of electromagnetic waves of different orders in Hall 1 and the actual channel impulse response is presented. It can be observed that the power difference between the first-order and second-order paths and the dominant path (i.e., the multipath with the highest power) is mostly within 30 dB. These two types of paths typically constitute the main contribution to channel energy and are key factors determining the overall channel characteristics. The model performs particularly well in predicting the first-order and second-order propagation components, effectively reproducing the power attenuation trend, indicating that the network effectively learns the dominant influence of environmental geometry and material properties on the propagation path. In contrast, the power of the third-order multipath is significantly lower than the first two orders, belonging to the weak energy tail component of the channel. Although its impact on the overall channel is small, the model still maintains basic trend consistency. The overall results show that the proposed model has extremely high accuracy in predicting key dominant multipaths (first-order and second-order), thus effectively determining the main propagation characteristics of the channel.

[0081] For example, Figure 6 and Figure 7 This paper presents a comparison between the actual multipath scattering results and the model predictions at a typical location. The multipath paths with continuous time delay and a regular decrease in power correspond to multipath clusters generated by different scatterers. It can be observed that within the same cluster, the multipath power exhibits a steady-state and slow decreasing trend with the incident angle, consistent with the physical expectations of the classical uniform scattering model. Notably, the model predictions also reproduce the gradual power change with angle in the corresponding angular domain, and the change pattern is consistent with the actual physical laws. The results indicate that the network can capture material-related scattering energy attenuation patterns and learn and encode the implicit directional response characteristics of different materials in the angular domain.

[0082] For example, Figure 8 The generalization results of the trained model in the Hall 2 scene are presented. The main structure of Hall 2 is consistent with that of Hall 1, but the positions of key scatterers within it have changed significantly, resulting in differences in the local propagation environment compared to the training scene. The results show that the model can still accurately reconstruct the main reflection and scattering paths under the new structural configuration, and the predicted multipath delay, angle, and power distribution maintain high consistency with the measured results. These results demonstrate that the model can not only capture the geometric and material features in static environments but also has the ability to adapt to changes in scatterer positions or adjustments in local structures, validating its effective generalization performance in dynamic or semi-dynamic scenes.

[0083] For example, Figure 9The generalization results of the trained model in the Hall 3 scenario are presented. Compared to Hall 1, the internal structure of Hall 3 has undergone significant adjustments, including changes in wall shape, scatterer layout, and the location of local openings, but its main constituent materials remain consistent with the training environment. Under these conditions of more drastic structural changes and greater spatial layout differences, the model can still accurately reconstruct the propagation paths of the main multipath clusters, and its predicted direction distribution and power variation trends are highly consistent with the measured results. This phenomenon indicates that the model can not only adapt to local geometric perturbations but also generalize to new scenarios with significant structural differences while maintaining similar material properties, verifying that the proposed method has good transferability and robustness in real-world complex architectural environments.

[0084] Therefore, this invention eliminates the need for manually constructing explicit 3D physical models, significantly reducing modeling complexity and improving efficiency. By directly processing and extracting features from measured point cloud data, it avoids the reliance on CAD models or mesh models found in traditional deterministic modeling methods, reducing the workload of manual modeling and scene reconstruction. This effectively lowers modeling costs and improves modeling efficiency in complex environments. It can integrate environmental geometry and material properties to improve the accuracy of electromagnetic propagation path modeling. Through neural networks, it jointly models point cloud geometry, surface morphology, and material properties, constructing implicit environmental feature representations. This allows electromagnetic propagation path modeling to simultaneously reflect the combined influence of spatial structure and material properties on propagation behavior, thereby improving the accuracy and physical consistency of multipath propagation modeling. Employing a hop-by-hop propagation prediction mechanism reduces the computational complexity of multipath path inference. By predicting electromagnetic propagation direction and power attenuation step-by-step through neural networks, it avoids exhaustive search of propagation paths, significantly reducing computational complexity while maintaining high simulation accuracy. This makes it suitable for efficient electromagnetic propagation path modeling in complex scenarios.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks, characterized in that, include: S1: Deploy a wireless channel measurement system and a synchronized 3D point cloud environment acquisition system to acquire wireless channel data and corresponding 3D spatial environment point cloud data of the target scene; S2: Preprocess the wireless channel data and point cloud data to obtain a physical environment feature dataset for neural network training; S3: Construct a three-dimensional environmental feature extraction model based on the physical environment feature dataset to generate an implicit environmental feature representation that reflects the relationship between environmental spatial distribution and material properties; S4: Based on the implicit environmental features, construct an electromagnetic propagation direction prediction model and a power prediction model, use a neural network to output multipath parameters hop by hop, predict the step-by-step evolution process of the electromagnetic propagation path, and output the electromagnetic propagation path-level characteristics.

2. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 1, characterized in that, In step S1, a wireless channel measurement system and a synchronized three-dimensional point cloud environment acquisition system are deployed, including: The target environment is divided into Region 1 and Region 2 based on differences in spatial structure. Among them, Region 1 and Region 2 have similar environmental characteristics but different physical structure layouts. The data in Region 1 is divided into training set and test set according to a preset ratio, and all the data in Region 2 is used as test set to evaluate the generalization ability of the model. During the actual test, the lidar and the signal receiver were deployed synchronously on the same mobile platform, while the signal transmitter was placed at a fixed location. At the start of the measurement, the signal receiver continuously receives and stores the vector detection signal, while the lidar synchronously collects and records the three-dimensional structural data of the environment in real time, realizing the time-synchronous storage of wireless channel data and three-dimensional spatial environment point cloud data.

3. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 2, characterized in that, In step S2, the wireless channel data and point cloud data are preprocessed to obtain a physical environment feature dataset for neural network training, including: The wireless channel data is noise-removed, and the channel transfer function and channel impulse response are obtained by performing a fast Fourier transform on the channel sounding signal. Multipath propagation parameters, including multipath delay, angle of arrival, received power and path loss, are extracted. The point cloud data is subjected to coordinate alignment, noise reduction filtering and downsampling processing, and the point cloud normal vector is estimated based on neighborhood features to construct an environmental feature display representation that includes geometric structure, surface morphology and material properties.

4. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 3, characterized in that, In step S3, a three-dimensional environmental feature extraction model is constructed based on the physical environment feature dataset to generate an implicit environmental feature representation that reflects the relationship between environmental spatial distribution and material properties, including: Fourier position encoding is performed on the point cloud coordinates to map the three-dimensional geometric information and material properties to a high-dimensional feature space; The PointNet++ network is used to extract multi-level features from the encoded point cloud. The PointNet++ network obtains local geometric features that can characterize the spatial structure at different scales by performing layer-by-layer sampling, neighborhood search and local feature aggregation on the point cloud. By constructing an implicit environmental feature expression that can simultaneously reflect the spatial structure, surface properties, and material properties of the environment through global feature fusion and nonlinear mapping structure, a well-trained electromagnetic environment feature extraction model can be obtained for subsequent electromagnetic propagation path modeling.

5. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 4, characterized in that, In step S4, an electromagnetic propagation direction prediction model is constructed based on the implicit environmental characteristics, including: An electromagnetic propagation direction prediction model based on a Transformer network is constructed. The Transformer encoder adopts a multi-head self-attention mechanism and a feedforward neural network, which includes 4 attention heads and 3 encoding layers, and outputs a three-dimensional direction vector. The Transformer network is trained using a dataset, with implicit environmental feature data as input; The loss function is constructed as the direction vector similarity between the true value and the predicted value, which is used to optimize the model performance. The direction vector similarity is measured by calculating the cosine similarity between the predicted direction vector and the actual direction vector. The Transformer network is trained using this loss function to obtain a trained Transformer model, which is used as an electromagnetic propagation direction prediction network to infer the next hop propagation direction.

6. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 5, characterized in that, In step S4, a power prediction model is constructed based on the implicit environmental features, including: The power prediction model is based on a multilayer perceptron network. The network contains four fully connected layers with layer sizes of 128, 256, 64, and 8, respectively. The network uses the ReLU activation function for nonlinear transformation, resulting in an 8-dimensional vector as the decay matrix for each hop.

7. The method for three-dimensional environment feature extraction and electromagnetic propagation path modeling based on neural networks according to claim 6, characterized in that, In step S4, the stepwise evolution process of the electromagnetic propagation path is predicted. The iteration process of the propagation path is limited by setting a maximum number of bounces, thus obtaining an electromagnetic propagation path prediction model.

Images