Channel knowledge map construction method and system based on neural ray tracing

Abstract

The application provides a channel knowledge map construction method and system based on neural ray tracing, comprising: collecting measured data of electromagnetic wave propagation by deploying channel detection equipment; preprocessing the measured data; training a deep learning model using the preprocessed data and a material library, so that the deep learning model can learn the complex nonlinear relationship between electromagnetic waves and the environment; through the trained neural network-based ray tracing framework, neural ray tracing is performed on each position after a given three-dimensional map and resolution, and relevant channel parameters are predicted according to the obtained multipath level channel information, to generate a channel knowledge map under the current environment and set appropriate values to visualize it. The application uses the fitting ability of neural networks for nonlinear relationships to calculate the interaction effect of rays and the environment, which can better predict the electromagnetic properties of materials in actual scenarios and has excellent migration ability.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a method and system for constructing a channel knowledge map based on neural ray tracing. Background Art

[0002] In the field of wireless communications, especially with the development of 5G and future communication technologies, Multiple-Input Multiple-Output (MIMO) technology has become key to improving data transmission rates and system capacity. The successful implementation of MIMO technology relies on the accurate acquisition of channel state information (CSI), which is crucial for beamforming and signal preprocessing to maximize signal gain and reduce interference. However, traditional channel acquisition methods mainly rely on the training process of pilot signals, which leads to increased time and energy overhead as the number of antennas increases, limiting the scalability and energy efficiency of the system.

[0003] In modern wireless communication systems, the concept of a channel knowledge map is widely used to describe and predict the propagation characteristics of wireless signals in a specific environment. This map not only encompasses the path loss and multipath effects of wireless signals, but also encompasses the electromagnetic characteristics of the environment, such as reflection, refraction, diffraction, and scattering. Channel knowledge maps are crucial for improving the performance of wireless communication systems, enabling accurate channel prediction, beamforming, network planning and optimization, and providing high-precision positioning services.

[0004] Traditional methods for constructing channel knowledge maps include interpolation, black-box neural networks, and traditional ray tracing. The advantage of interpolation is its simplicity and ease, enabling rapid generation of continuous channel knowledge maps. However, its disadvantage is its dependence on the quality and density of measured data, and prediction accuracy decreases when data is sparse or the environment is highly variable. The advantage of black-box neural networks is their ability to capture the complex relationship between channel characteristics and environmental features, making them suitable for modeling nonlinear and high-dimensional data. However, this approach suffers from poor model interpretability and the requirement for large amounts of training data. Furthermore, traditional methods for constructing channel knowledge maps face numerous challenges. First, these methods typically rely on large-scale field measurements, which are time-consuming, labor-intensive, and costly. Furthermore, as the environment changes, such as building renovations and the emergence of new obstacles, traditional channel knowledge maps require frequent updates, further increasing the complexity and cost of maintenance. Furthermore, traditional channel modeling methods often struggle to adapt to dynamically changing environments, resulting in decreased channel prediction accuracy.

[0005] Ray tracing is the most accurate method for channel knowledge maps. Ray tracing is a traditional method used in wireless communications to simulate the propagation of electromagnetic waves in complex environments. It predicts the characteristics of wireless channels by simulating the propagation paths of electromagnetic waves, including direct radiation, reflection, refraction, and scattering. This method provides a channel model that closely resembles the real world, and has therefore been widely used in the planning, optimization, and performance evaluation of wireless communication systems. The core of ray tracing technology lies in constructing a three-dimensional model that incorporates the geometry and material properties of the environment. This model is then used to calculate the interactions of electromagnetic waves with various objects in the environment. These interactions include reflection, refraction, and scattering of the waves when encountering obstacles. Each interaction affects the amplitude, phase, and polarization of the waves. By tracing all possible paths from the transmitter to the receiver, the response of the entire channel can be synthesized, providing important information for the design and analysis of wireless communication systems.

[0006] Although ray tracing technology can theoretically provide highly accurate channel models, it still faces several challenges and limitations in practical applications. First, traditional ray tracing techniques often rely on idealized physical models to simulate the interaction between electromagnetic waves and the environment. These models may not fully account for the complexity and variability of real-world environments, resulting in deviations from actual simulation results. Accurate ray tracing simulations require detailed environmental parameters, including the geometry and electromagnetic properties of objects. In practice, acquiring these parameters is time-consuming and expensive, especially for large or complex environments. Furthermore, traditional ray tracing models are often optimized for specific environments and lack the ability to migrate between them. This means that complex simulations and parameter adjustments may need to be performed from scratch in a new environment. The real-world communication environment is constantly changing, and traditional ray tracing techniques perform poorly in adapting to these changes. Any changes in the environment, such as the movement or addition of objects or the introduction of new materials, can affect channel characteristics, requiring the model to quickly adapt to these changes.

[0007] With the development of deep learning technology, neural networks have been introduced into wireless channel modeling to address the limitations of traditional methods. Neural networks can learn complex nonlinear relationships from data, offering new possibilities for channel modeling. In particular, Neural Radiance Field (NeRF) technology has achieved breakthroughs in computer vision, demonstrating the ability to synthesize novel perspectives from sparse samples. This provides a new perspective for wireless channel modeling and prediction.

[0008] Therefore, inspired by the NeRF concept, the present invention proposes a channel knowledge map construction method and system for neural ray tracing. Summary of the Invention

[0009] In view of the defects in the prior art, the purpose of the present invention is to provide a channel knowledge map construction method and system based on neural ray tracing.

[0010] According to the present invention, a method for constructing a channel knowledge map based on neural ray tracing is provided, comprising:

[0011] Step S1: Deploy channel detection equipment in various typical environments to collect measured data on electromagnetic wave propagation;

[0012] Step S2: pre-processing the measured data;

[0013] Step S3: using the pre-processed data and material library to train a deep learning model, so that the deep learning model can learn the complex nonlinear relationship between the interaction between electromagnetic waves and the environment;

[0014] Step S4: After a given three-dimensional map and resolution, neural ray tracing is performed on each position using a trained neural network-based ray tracing framework. Based on the obtained multipath-level channel information, relevant channel parameters are predicted to generate a channel knowledge map in the current environment and set appropriate values ​​to visualize it.

[0015] Preferably, the measured data is the basis for the model to learn the electromagnetic wave propagation characteristics in the actual environment, including detailed channel state information;

[0016] The channel state information includes a channel impulse response;

[0017] The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training;

[0018] The relevant channel parameters include channel gain, delay spread, and angle spread.

[0019] Preferably, in step S3, signal propagation is modeled using a trained neural network, and the steps include:

[0020] Step S3.1: Using a traditional ray tracing algorithm, find all reachable paths and all interaction points between the ray and the environment;

[0021] Step S3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain the material information embedding vector;

[0022] Step S3.3: Input the material embedding vector and incident wave information into the shared layer and reflection / scattering loss prediction network to obtain the corresponding loss of this interaction;

[0023] Step S3.4: Sum up all interaction losses of each path to obtain the total path loss and the received power at the receiving end.

[0024] Preferably, the step S4 includes predicting a multipath narrowband channel model, wherein the multipath narrowband channel model is as follows:

[0025]

[0026] Where L represents the number of propagation paths, f represents the carrier frequency, and n r With n t Respectively represent the number of antenna units at the receiving end and the transmitting end, α l ∈C and τ l ∈R represent the amplitude loss and propagation delay of the lth path respectively;

[0027] According to the number of interactions between the lth path and the environment, the amplitude of the electric field gradually decreases from the transmitter to the receiver due to the interaction with the object. Let E l,i represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows:

[0028]

[0029] Among them, T l,i (·) represents the transfer function of the electric field of the ith interaction to the lth path, I l is the total number of interactions in the lth path.

[0030] Preferably, the method further includes a model evaluation and optimization step, wherein the prediction performance of the model is evaluated by comparing the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map, and necessary optimization is performed;

[0031] The optimization includes adjusting the model structure, training strategy and loss function;

[0032] The optimization includes expressing the training objective of the neural network as an optimization problem of the following formula:

[0033]

[0034] Among them, ε represents the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

[0035] According to the present invention, a channel knowledge map construction system based on neural ray tracing is provided, comprising:

[0036] Module M1: Collect measured data on electromagnetic wave propagation by deploying channel detection equipment in various typical environments;

[0037] Module M2: pre-processing the measured data;

[0038] Module M3: Using the pre-processed data and material library, training a deep learning model to enable the deep learning model to learn the complex nonlinear relationship between electromagnetic waves and the environment;

[0039] Module M4: Using a trained neural network-based ray tracing framework, given a 3D map and resolution, neural ray tracing is performed at each location. Based on the obtained multipath-level channel information, relevant channel parameters are predicted, and a channel knowledge map is generated for the current environment and visualized by setting appropriate values.

[0040] Preferably, the measured data is the basis for the model to learn the electromagnetic wave propagation characteristics in the actual environment, including detailed channel state information;

[0041] The channel state information includes a channel impulse response;

[0042] The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training;

[0043] The relevant channel parameters include channel gain, delay spread, and angle spread.

[0044] Preferably, in the module M3, signal propagation is modeled using a trained neural network, and the steps include:

[0045] Module M3.1: Use traditional ray tracing algorithms to find all reachable paths and all interaction points between rays and the environment;

[0046] Module M3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain the material information embedding vector;

[0047] Module M3.3: Input the material embedding vector and incident wave information into the shared layer and reflection / scattering loss prediction network to obtain the corresponding loss of this interaction;

[0048] Module M3.4: Summarize all interaction losses on each path to obtain the total path loss and the received power at the receiving end.

[0049] Preferably, the module M4 includes predicting a multipath narrowband channel model, wherein the multipath narrowband channel model is as follows:

[0050]

[0051] Where L represents the number of propagation paths, f represents the carrier frequency, and n r With n t Respectively represent the number of antenna units at the receiving end and the transmitting end, α l ∈C and τ l ∈R represent the amplitude loss and propagation delay of the lth path respectively;

[0052] According to the number of interactions between the lth path and the environment, the amplitude of the electric field gradually decreases from the transmitter to the receiver due to the interaction with the object. Let E l,i represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows:

[0053]

[0054] Among them, T l,i (·) represents the transfer function of the electric field of the ith interaction to the lth path, I l is the total number of interactions in the lth path.

[0055] Preferably, a model evaluation and optimization module is also included, which evaluates the prediction performance of the model by comparing the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map and performs necessary optimization;

[0056] The optimization includes adjusting the model structure, training strategy and loss function;

[0057] The optimization includes expressing the training objective of the neural network as an optimization problem of the following formula:

[0058]

[0059] Among them, ε represents the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. Based on actual channel measurement data, this method leverages the excellent fitting capabilities of neural networks for nonlinear relationships to calculate the interaction between rays and the environment, including reflection and scattering. Compared to traditional ray tracing solutions, this method can better predict the electromagnetic properties of materials in real-world scenarios and has excellent transferability.

[0062] 2. The neural ray tracing technology solution of the present invention effectively improves the accuracy and environmental adaptability of wireless channel modeling, reduces dependence on field measurements, and saves time and costs.

[0063] 3. The model of the present invention can quickly adapt to environmental changes and improve the reliability of channel prediction.

[0064] 4. The neural ray tracing method of the present invention enhances the model's ability to migrate between different environments, providing a powerful tool for the design, optimization and performance evaluation of wireless communication systems, and has important theoretical and practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] 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:

[0066] Figure 1 Correlation between digital twins and real-world scenarios for ray tracing;

[0067] Figure 2 The electric field changes as the signal interacts with the object along the ray tracing path.

[0068] Figure 3 The network architecture used by neural ray tracing to predict the outgoing electric field;

[0069] Figure 4 Taking concrete as an example, the prediction effect of neural ray tracing on reflection loss is demonstrated;

[0070] Figure 5 Taking concrete as an example, the prediction effect of neural ray tracing on scattering loss is demonstrated;

[0071] Figure 6 Comparison of the channel knowledge map constructed for neural ray tracing and the real channel knowledge map;

[0072] Figure 7 NMSE of the received power channel knowledge map constructed by different methods at different sampling rates;

[0073] Figure 8 The received power prediction effects of two neural ray tracing network structures for different types of paths. DETAILED DESCRIPTION

[0074] 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.

[0075] The present invention uses neural networks to learn the complex nonlinear relationship between electromagnetic waves and the environment to achieve high-precision channel knowledge map construction. By learning the electromagnetic properties of different materials in the environment, it can not only improve the accuracy of model predictions, but also enhance the model's ability to migrate between different environments and its adaptability to dynamically changing environments.

[0076] According to the present invention, a method for constructing a channel knowledge map based on neural ray tracing is provided, comprising:

[0077] Step S1: Deploy channel detection equipment in a variety of typical environments to collect measured data on electromagnetic wave propagation. The measured data includes detailed channel state information, such as channel impulse response, which is the basis for the model to learn the electromagnetic wave propagation characteristics in the actual environment.

[0078] Step S2: Preprocess the measured data. The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training.

[0079] Step S3: Using the preprocessed data and material library, a deep learning model is trained to learn the complex nonlinear relationship between electromagnetic waves and the environment. During the model training process, the present invention uses advanced optimization algorithms and strategies to ensure that the model can effectively learn from the data, learn the electromagnetic properties of different materials using an end-to-end training framework, and accurately predict channel characteristics. Using the trained neural network, we model signal propagation in the following sub-steps:

[0080] Step S3.1: Use the traditional ray tracing algorithm to find all reachable paths and all interaction points between the ray and the environment.

[0081] Step S3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain a material information embedding vector.

[0082] Step S3.3: Input the material embedding vector and incident wave information into the shared layer and reflection / scattering loss prediction network to obtain the corresponding loss of this interaction.

[0083] Step S3.4: Sum up all interaction losses of each path to obtain the total path loss and the received power at the receiving end.

[0084] Step S4: Using a trained neural network-based ray tracing framework, given a 3D map and resolution, neural ray tracing is performed at each location. Based on the obtained multipath-level channel information, channel gain, delay spread, angle spread, and other related channel parameters are predicted. A channel knowledge map for the current environment is generated and visualized by setting appropriate values. Step S4 includes predicting the wireless channel using the following formula:

[0085]

[0086] in, represents a neural network that encodes the electromagnetic characteristics of the environment, where Θ ε represents the trainable network parameters, represents a neural network that computes interaction effects, represents the trainable network parameters, Represents all geometric information in the environment, Indicates related hardware devices, Indicates the reachable path information between the sender and receiver found by the ray tracing algorithm.

[0087] The step S4 includes predicting the multipath narrowband channel model, wherein the multipath narrowband channel model is as follows:

[0088]

[0089] Where L represents the number of propagation paths, f represents the carrier frequency, and n r With n t Respectively represent the number of antenna units at the receiving end and the transmitting end, α l ∈C and τ l ∈R represent the amplitude loss and propagation delay of the lth path respectively.

[0090] According to the number of interactions between the lth path and the environment, the amplitude of the electric field gradually decreases from the transmitter to the receiver due to the interaction with the object. Let E l,i represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows:

[0091]

[0092] Among them, T l,i (·) represents the transfer function of the electric field of the ith interaction to the lth path, I l is the total number of interactions in the lth path.

[0093] In addition, the present invention compares the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map to evaluate the model's predictive performance and perform necessary optimization. The optimization includes adjusting the model structure, training strategy, and loss function to further improve the accuracy and robustness of the model. This step is performed iteratively to ensure that the model maintains optimal performance under various environments and conditions. The optimization includes expressing the training objective of the neural network as an optimization problem as follows:

[0094]

[0095] Among them, ε represents the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

[0096] The present invention aims to address the challenges faced by existing ray tracing technology in wireless channel modeling, especially the problem of deviation between simulation results and actual conditions due to idealized physical models in practical applications. Existing technologies often fail to fully capture the complexity and variability of the actual environment when simulating the interaction between electromagnetic waves and the environment, and the accurate acquisition of environmental parameters is both time-consuming and expensive. In addition, traditional models lack the ability to migrate between different environments and have poor adaptability to dynamic environmental changes, which limits the applicability and accuracy of the model in diverse and changing environments. Therefore, the present invention seeks to propose a new wireless channel modeling method to improve the accuracy, adaptability and migration capability of the model, so that it can better reflect the electromagnetic wave propagation characteristics in the actual communication environment and effectively respond to the challenges brought about by environmental changes.

[0097] Furthermore, the theoretical basis and implementation process of this solution are described in detail. In actual scenarios, the channel is composed of the communication environment ε, which includes the location, orientation, shape and electromagnetic characteristics of the base station, users and other objects, and the signal propagation law g(·) and related hardware equipment. (such as the antenna pattern of the transmitting and receiving ends, etc.). Therefore, the communication channel can be expressed as:

[0098]

[0099] The ray tracing method is a form of expression of the signal propagation law g(·). It uses the geometric and electromagnetic characteristics of the environment in various interactive modes to find feasible propagation paths to simulate the channel between the transmitter and receiver. The interactive modes include reflection, diffraction, and scattering. Ray tracing can calculate the loss, delay, and angle of each path. The relevant parameters of these paths are used to extract channel information. Further, ray tracing is divided into two parts: (1) Path tracing Calculate all reachable paths between the transmitter and receiver, where is all the geometric information in the environment. (2) Loss calculation It is used to model the loss of electromagnetic waves on a given path and integrate the information of all paths into the transceiver channel.

[0100] However, obtaining accurate electromagnetic property information ε of objects in an environment is extremely difficult. First, traditional methods for measuring electromagnetic properties require specialized equipment and environments (e.g., a darkroom). Therefore, measuring the electromagnetic properties of materials in real-world scenarios is often difficult to achieve. Furthermore, currently used electromagnetic parameters such as conductivity and magnetic permeability are often insufficient to accurately characterize a material's electromagnetic properties (a material's electrical properties are also dependent on ambient temperature, humidity, and other factors).

[0101] At the same time, Traditional ray tracing often relies on readily available theoretical models for modeling, such as Fresnel's law to calculate reflection losses. However, these theoretical formulas often fail to accurately describe the effects of more complex interactions on radio wave propagation, such as scattering and diffraction. Furthermore, such calculations are further complicated when the electromagnetic properties of materials are highly nonlinear.

[0102] To address the above challenges, the present invention uses neural networks to model the electromagnetic properties of various materials and the effects of various interaction modes on signal propagation. represents a neural network that encodes the electromagnetic properties of the environment, where Θ ε Represents the trainable network parameters. A neural network representing a computational interaction effect can predict the corresponding loss based on the electromagnetic properties of the corresponding material (represented by the material embedding vector). Based on the electromagnetic properties encoded by the neural network and the modeling of the corresponding interaction effects, the wireless channel can be predicted:

[0103]

[0104] Therefore, the training objective of the neural network of the present invention can be expressed as the following optimization problem:

[0105]

[0106] Where ε represents the communication scenario, Represents a sampling channel in this communication scenario, Represents the difference between the real channel and the predicted channel.

[0107] Specifically, the present invention considers a multipath narrowband channel model, and the channel can be written as:

[0108]

[0109] Where L represents the number of propagation paths, f represents the carrier frequency, and n r With n t Represents the number of antenna units at the receiving end and the transmitting end, α l ∈C and τ l ∈R represent the amplitude loss and propagation delay of the lth path respectively. The present invention can be achieved by the electric field E on the lth path. l Iteration to characterize the amplitude loss α of the path l .exist Figure 2 In this paper, it is assumed that the kth path has two interactions with the environment. From the transmitting end to the receiving end, the amplitude of the electric field gradually decreases due to the interaction with the object. The electric field is as follows:

[0110]

[0111] in Represent two orthogonal polarization directions. s ,E p ∈C represent the amplitude components of the electric field E in these two directions. Let E l,i represents the amplitude of the electric field of the incident wave in the i-th interaction, then the electric field of the i+1-th interaction (or the electric field strength at the receiver end) is:

[0112] E l,i+1 =T l,i (E l,i )

[0113] (6) T l,i (·) represents the transfer function of the electric field of the ith interaction to the lth path. Therefore, the electric field at the receiver is expressed as:

[0114]

[0115] Among them I l is the total number of interactions in the lth path. As shown above, the total transfer function of the electric field in the lth path can be written as a cascade of several interaction transfer functions. Specifically, T l,i Depending on the type of interaction (reflection, scattering, etc.), the electromagnetic properties of the material at the interaction point and the geometric relationship of the electromagnetic wave path relative to the interaction surface. Come to E l,i Perform the calculation, namely:

[0116]

[0117] Substituting (6)(7)(8) into (4), the present invention can train the network by solving the following optimization problem:

[0118]

[0119] Figure 3 Shown In every interaction, E l,i Predicted output electric field E l,i+1 The neural network architecture used. The network for predicting the shared intermediate layer and reflection / scattering loss is based on encoding the electromagnetic properties of the material and then further calculating the loss caused by this interaction on the radio wave. The specific training is done through equation (9), that is, it is hoped that the electric field strength received by the final receiver predicted by the network is consistent with the actual situation. Figure 3 As shown, the electric field prediction for each interaction can be calculated as follows:

[0120] The first step is to select the one-hot encoding vector of the corresponding material based on the location of the interaction point.

[0121] In the second step, based on the auxiliary information (carrier frequency, environmental climate, etc.), the material information embedding network is used to obtain the material information embedding vector of the material in the current communication environment.

[0122] The third step is to combine the incident direction, incident polarization mode, and interaction surface normal vector to obtain the intermediate vector through the shared intermediate layer of scattering and reflection.

[0123] The fourth step is to determine whether there is a LoS path between the interaction point and the receiving end to decide whether to use the reflection loss prediction network or the scattering loss prediction loss, and calculate the output electric field (if it is scattering, the offset angle of the scattering direction compared to the reflection direction is added as input).

[0124] By using the neural network to calculate the attenuation of each interaction electric field, combined with By searching all reachable paths, we can use (2) to obtain the channel predicted by neural ray tracing. In actual scenarios, the present invention can obtain the actual channel through measurement By minimizing the difference between the predicted channel and the actual channel, the neural network of the present invention can learn end-to-end.

[0125] Figure 4Using concrete as an example, the performance of the neural network model developed in this study in predicting the return loss of concrete for both parallel and perpendicular polarizations at different incident angles is demonstrated. The model, trained and predicted using a complex network architecture combining a multilayer perceptron (MLP) and a Transformer, operates at a 3.6 GHz carrier frequency and a 15% sampling rate. Comparing the model's predictions with the experimentally measured true values ​​clearly demonstrates the high accuracy of the neural network model in simulating the electromagnetic wave reflection characteristics of concrete. Under parallel polarization, the model's predictions and true values ​​maintain good agreement across a wide range of incident angles. However, near the Brewster angle (approximately 67 degrees), a certain degree of deviation between the model's predictions and the true values ​​occurs. This phenomenon is likely due to the drastic changes in the electromagnetic properties of concrete at the Brewster angle, resulting in a decrease in the model's prediction accuracy at this specific angle. Furthermore, the comparison of the predicted and true values ​​further validates the model's stability and reliability in simulating the reflection characteristics of concrete. Although there are certain prediction errors at specific angles, overall the model's prediction performance is quite outstanding, which has important theoretical and practical significance for the study of the propagation characteristics of electromagnetic waves in various materials and related engineering applications.

[0126] Figure 5 Using concrete as an example, the neural network model developed in this study demonstrates its performance in predicting the scattering loss of concrete at various scattering offset angles at an incident angle of 45 degrees. This simulation, also conducted at a 3.6 GHz carrier frequency and a 15% sampling rate, combines the strengths of a multilayer perceptron (MLP) and a transformer to form a complex network architecture. This network architecture, trained with extensive data, aims to accurately simulate the scattering properties of concrete. The figure clearly shows that the model-predicted scattering loss (grey points) and the measured scattering loss (black points) show high agreement at most scattering offset angles, demonstrating that the model effectively captures the scattering behavior of concrete at specific incident angles. As the scattering offset angle increases, the scattered wave electric field intensity gradually decreases, a trend accurately reflected in the model's predictions, further confirming the model's effectiveness in simulating the scattering properties of concrete. In summary, the neural network model proposed in this study demonstrates high accuracy and reliability in predicting the scattering loss of concrete. This not only provides strong theoretical support for the analysis of electromagnetic properties of various materials but also has important practical implications for predicting electromagnetic wave propagation in related engineering applications. Future research can further optimize the model on this basis to improve its applicability and prediction accuracy under a wider range of conditions.

[0127] Figure 6Two sub-figures (a) and (b) illustrate channel knowledge maps of user received power in a building environment, constructed using the neural ray tracing technology proposed in this study, at a sampling rate of 15%. These maps reflect the signal strength that users at different locations can expect to receive, given a specific transmitter location. Figure (a) represents the channel knowledge map constructed using neural ray tracing, while Figure (b) represents the actual channel knowledge map obtained through actual measurements. Comparing the two figures, we can observe significant similarity between the maps constructed using neural ray tracing and the actual measurement results. This similarity demonstrates that neural ray tracing can effectively simulate and predict the propagation characteristics of electromagnetic waves in building environments, including phenomena such as direct radiation, reflection, and obstruction. In Figures (a) and (b), black areas represent regions of low received signal power, typically due to obstruction by building structures. Regions with a gradient from gray to white represent regions of gradually increasing received signal power. This color coding allows us to intuitively demonstrate that neural ray tracing not only identifies areas of signal obstruction but also accurately depicts the path and intensity changes of signals due to reflection and diffraction. Furthermore, the construction of this channel knowledge map is of great significance for the design and optimization of wireless communication systems. It can help engineers better understand the propagation behavior of signals in real environments, thereby making more reasonable decisions in network planning, base station deployment, and signal coverage optimization.

[0128] Figure 7The changing trends of the normalized root mean square error (NMSE) of the four received power channel knowledge map construction methods at different sampling rates are shown. As the sampling rate increases from 0.05 to 0.30, the NMSE of all methods shows a downward trend, which shows that increasing the sampling rate helps to reduce the prediction error and improve the accuracy of channel knowledge map construction. Under low sampling rate (0.05) conditions, RadioUNet shows the best received power prediction effect. This is because RadioUNet maintains a certain spatial understanding ability during the training process. Even if there is no measurement data or the measurement data is small, it still has a good channel knowledge map construction ability. Under low sampling rate conditions, the NeRT method based on MLP+Transformer shows a lower NMSE. Its advantage becomes more obvious as the sampling rate increases, and finally maintains the lowest NMSE among all methods, indicating that this method has good performance and generalization ability when dealing with channel knowledge map construction tasks. Furthermore, the MLP+Transformer-based NeRT method demonstrates significant superiority over the MLP-based NeRT method at lower sampling rates, but this superiority diminishes at higher sampling rates. This suggests that the addition of the attention mechanism significantly aids the network in constructing channel knowledge maps in the presence of insufficient samples. The NMSE of the RadioUNet and MLP-based NeRT methods also decreases with increasing sampling rates, demonstrating a certain degree of prediction accuracy, but their overall performance is slightly inferior to that of the MLP+Transformer-based NeRT method. The Kriging Interpolation method also achieves a low NMSE at high sampling rates, but its error is relatively high at low sampling rates. Overall, the MLP+Transformer-based NeRT method maintains a low NMSE across different sampling rates, demonstrating its superior performance in channel knowledge map construction. This result provides important guidance for selecting appropriate channel knowledge map construction methods in practical applications. In particular, the MLP+Transformer-based NeRT method may be a more reliable choice when sampling resources are limited.

[0129] Figure 8This paper presents the Normalized Mean Square Error (NMSE) performance of two network architectures in predicting various path types, which are distinguished by the number of reflections and scatterings an electromagnetic wave experiences during propagation. The figure includes models using only a multilayer perceptron (MLP) and a combination of an MLP and a Transformer (MLP+Transformer). Each model considers different sampling rates (10%, 20%, and 30%). As can be seen from the figure, for paths with a single reflection, all models achieve relatively low NMSE, indicating that the models have good prediction performance for simple reflection paths. As the number of reflections increases, especially for paths with three and four reflections, the NMSE of the MLP model increases significantly, indicating a decline in performance for complex reflection paths. In contrast, the MLP+Transformer model exhibits lower NMSE for all path types, with a particularly significant performance advantage for paths with three and four reflections. This may be because the Transformer architecture effectively captures long-range dependencies, resulting in better performance in predicting complex paths. For scattered paths, the MLP model exhibits a relatively high NMSE, while the MLP+Transformer model significantly reduces this error. In particular, at a sampling rate of 30%, its NMSE approaches zero, demonstrating superior performance for scattered paths. Overall, the combined MLP and Transformer model demonstrates superior prediction performance for a variety of complex paths, especially at high sampling rates. This result highlights the importance of considering model complexity and sampling rate in channel modeling, as well as the potential of the MLP+Transformer model to improve prediction accuracy.

[0130] The present invention also provides a channel knowledge map construction system based on neural ray tracing. The channel knowledge map construction system based on neural ray tracing can be implemented by executing the process steps of the channel knowledge map construction method based on neural ray tracing. That is, those skilled in the art can understand the channel knowledge map construction method based on neural ray tracing as a preferred implementation of the channel knowledge map construction system based on neural ray tracing.

[0131] According to the present invention, a channel knowledge map construction system based on neural ray tracing is provided, comprising:

[0132] Module M1: Deploy channel detection equipment in various typical environments to collect measured data on electromagnetic wave propagation. This data forms the basis for the model to learn the characteristics of electromagnetic wave propagation in real environments, including detailed channel state information, including the channel impulse response.

[0133] Module M2: Preprocessing the measured data. The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training.

[0134] Module M3: Use the pre-processed data and material library to train the deep learning model so that the deep learning model can learn the complex nonlinear relationship between the interaction between electromagnetic waves and the environment. In the module M3, the signal propagation is modeled through the trained neural network, including: Module M3.1: Use the traditional ray tracing algorithm to find all reachable paths and all interaction points between the rays and the environment. Module M3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain the material information embedding vector. Module M3.3: The material embedding vector and the incident wave information are input into the shared layer and the reflection / scattering loss prediction network to obtain the corresponding loss of this interaction. Module M3.4: All interaction losses of each path are summarized to obtain the total path loss and the receiving power at the receiving end.

[0135] Module M4: Using a trained neural network-based ray tracing framework, given a 3D map and resolution, neural ray tracing is performed at each location. Based on the obtained multipath-level channel information, relevant channel parameters are predicted, a channel knowledge map for the current environment is generated, and appropriate values ​​are set to visualize it. The relevant channel parameters include channel gain, delay spread, and angle spread. Module M4 includes predicting the multipath narrowband channel model, which is as follows:

[0136]

[0137] Where L represents the number of propagation paths, f represents the carrier frequency, and n r With n t Respectively represent the number of antenna units at the receiving end and the transmitting end, α l ∈C and τ l ∈R represent the amplitude loss and propagation delay of the lth path respectively. According to the number of interactions between the lth path and the environment, the amplitude of the electric field gradually decreases from the transmitter to the receiver due to the interaction with the object, let E l,i represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows:

[0138]

[0139] Among them, T l,i (·) represents the transfer function of the electric field of the ith interaction to the lth path, I l is the total number of interactions in the lth path.

[0140] The present invention also includes a model evaluation and optimization step, which compares the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map to evaluate the model's predictive performance and perform necessary optimization. The optimization includes adjusting the model structure, training strategy, and loss function. The optimization includes expressing the training objective of the neural network as an optimization problem of the following formula:

[0141]

[0142] Among them, ε represents the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

[0143] 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.

[0144] 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 channel knowledge map construction method based on neural ray tracing, characterized in that: include: Step S1: collecting measured data of electromagnetic wave propagation; Step S2: pre-processing the measured data; Step S3: using the pre-processed data and material library to train a deep learning model, so that the deep learning model can learn the complex nonlinear relationship between the interaction between electromagnetic waves and the environment; Step S4: Using the trained neural network-based ray tracing framework, given a 3D map and resolution, neural ray tracing is performed at each location. Based on the obtained multipath-level channel information, relevant channel parameters are predicted to generate and visualize the channel knowledge map in the current environment. In step S3, signal propagation is modeled using a trained neural network, and the steps include: Step S3.1: Using a traditional ray tracing algorithm, find all reachable paths and all interaction points between the ray and the environment; Step S3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain the material information embedding vector; Step S3.3: Input the material embedding vector and incident wave information into the shared layer and reflection / scattering loss prediction network to obtain the corresponding loss of this interaction; Step S3.4: Sum up all interaction losses of each path to obtain the total path loss and the received power at the receiving end; The step S4 includes predicting the multipath narrowband channel model, wherein the multipath narrowband channel model is as follows: in, represents the number of propagation paths, Indicates the carrier frequency, and Respectively represent the number of antenna units at the receiving end and the transmitting end, and Respectively represent Amplitude loss and propagation delay of each path; According to The number of interactions between the path and the environment, from the transmitter to the receiver, the amplitude of the electric field gradually decreases due to the interaction with the object, represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows: in, Representative i The second interaction The transfer function of the electric field along the path, It is The total number of interactions along the path.

2. The method for constructing a channel knowledge map based on neural ray tracing according to claim 1, characterized in that: The measured data is the basis for the model to learn the electromagnetic wave propagation characteristics in the actual environment, including detailed channel state information; The channel state information includes a channel impulse response; The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training; The relevant channel parameters include channel gain, delay spread, and angle spread.

3. The method for constructing a channel knowledge map based on neural ray tracing according to claim 1, characterized in that: It also includes model evaluation and optimization steps, which evaluate the model's prediction performance and optimize it by comparing the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map; The optimization includes adjusting the model structure, training strategy and loss function; The optimization includes expressing the training objective of the neural network as an optimization problem of the following formula: in, Indicates the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

4. A channel knowledge map construction system based on neural ray tracing, characterized in that: include: Module M1: Collect measured data on electromagnetic wave propagation; Module M2: pre-processing the measured data; Module M3: Using the pre-processed data and material library, training a deep learning model to enable the deep learning model to learn the complex nonlinear relationship between electromagnetic waves and the environment; Module M4: Using a trained neural network-based ray tracing framework, given a 3D map and resolution, neural ray tracing is performed at each location. Based on the obtained multipath-level channel information, relevant channel parameters are predicted to generate and visualize a channel knowledge map for the current environment. In the module M3, signal propagation is modeled using a trained neural network, including: Module M3.1: Use traditional ray tracing algorithms to find all reachable paths and all interaction points between rays and the environment; Module M3.2: For each interaction of each path, the material number and auxiliary information are input into the material information embedding network to encode the electromagnetic properties of the material and obtain the material information embedding vector; Module M3.3: Input the material embedding vector and incident wave information into the shared layer and reflection / scattering loss prediction network to obtain the corresponding loss of this interaction; Module M3.4: Summarize all interaction losses on each path to obtain the total path loss and the received power at the receiving end; The module M4 includes predicting a multipath narrowband channel model, wherein the multipath narrowband channel model is as follows: in, represents the number of propagation paths, Indicates the carrier frequency, and Respectively represent the number of antenna units at the receiving end and the transmitting end, and Respectively represent Amplitude loss and propagation delay of each path; According to The number of interactions between the path and the environment, from the transmitter to the receiver, the amplitude of the electric field gradually decreases due to the interaction with the object, represents the amplitude of the electric field of the i-th interactive incident wave, and the electric field amplitude at the receiving end is as follows: in, Representative i The second interaction The transfer function of the electric field along the path, It is The total number of interactions along the path.

5. The channel knowledge map construction system based on neural ray tracing according to claim 4 is characterized in that: The measured data is the basis for the model to learn the electromagnetic wave propagation characteristics in the actual environment, including detailed channel state information; The channel state information includes a channel impulse response; The preprocessing includes data cleaning, normalization, and feature extraction to ensure data quality and prepare for subsequent model training; The relevant channel parameters include channel gain, delay spread, and angle spread.

6. The channel knowledge map construction system based on neural ray tracing according to claim 4 is characterized in that: It also includes a model evaluation and optimization module, which evaluates the model's prediction performance and optimizes it by comparing the channel knowledge map constructed by neural ray tracing with the actual channel knowledge map; The optimization includes adjusting the model structure, training strategy and loss function; The optimization includes expressing the training objective of the neural network as an optimization problem of the following formula: in, Indicates the communication scenario, Represents a sampling channel in this communication scenario, It represents the difference between the real channel and the predicted channel.

Images