Terahertz environment semantic map construction method based on physical information flow matching model

Abstract

The application discloses a terahertz environment semantic map construction method based on a physical information flow matching model, and relates to the technical field of communication.The method comprises the following steps: constructing a sparse directional beam received signal strength dataset of a real scene collection, a real omnidirectional radio map and a training set of a real physical environment; constructing a flow matching model with a built-in three-class propagation region dynamic judgment module, taking the sparse data as a condition, integrating a Helmholtz equation constraint and a regionalized power boundary constraint as double physical constraints, synchronously generating a radio map and a physical environment by adopting a double-output U-Net architecture, and training through a non-classifier guiding mechanism; in actual application, sparse directional beam data is collected to input the trained model, and an omnidirectional radio map and a physical environment of a current scene are generated by reasoning.The application solves the problems of sparse data and poor physical consistency in terahertz communication, reduces the hardware and time cost, and improves the map construction precision and generalization ability.

Description

Technical Field

[0001] This invention relates to communication technology, specifically to a joint sensing technology for wireless environment semantics and physical layout assisted by multi-beam sparse sensing and flow matching model based on terahertz communication signals. Background Technology

[0002] Terahertz communication, with its abundant spectrum resources and ultra-high data rates, is widely recognized as a core enabling technology for 6G wireless networks, supporting immersive communication, integrated air-space-ground communication, and other scenarios. However, the strong directionality of terahertz signals due to their ultra-short wavelengths leads to severe propagation limitations. The extremely weak diffraction capability of these signals makes them unable to bypass obstacles, making them easily blocked by walls and pedestrians. This, coupled with molecular absorption and path loss, further reduces their coverage area compared to low-frequency signals, posing a significant threat to the reliability of terahertz communication. To address this, Terahertz Information Sensing Integration (THz ISAC) technology has emerged. This technology multiplexes communication signals to achieve high-resolution environmental sensing and provides obstacle information without additional hardware, supporting environment-aware beam alignment and network optimization, ultimately improving communication stability in complex scenarios.

[0003] To unleash the potential of THz ISAC, an accurate semantic map of the radio propagation environment is indispensable. A terahertz radio environment semantic map comprises two parts: a radio map and a physical environment map. The radio map comprehensively describes the characteristics of the radio propagation environment because it captures the spatial distribution and directional characteristics of signal strength, reflecting the modulation effects of obstacles such as obstruction and reflection. Furthermore, radio maps for different scenarios can present differentiated signal shadow areas, characteristics directly related to the location and shape of obstacles. The physical environment map, on the other hand, provides a precise environmental reconstruction. With accurate environmental knowledge, terahertz communication can be enhanced through methods such as rapid beam alignment, user tracking, and interference minimization. The radio map and physical environment map work together to form a complete terahertz radio environment semantic representation. Based on this semantic representation, terahertz communication systems can achieve functions such as active beam avoidance, adaptive network reconstruction, and communication resource allocation, providing core support for the robustness and high quality of terahertz communication.

[0004] Currently, the collaborative construction of radio maps and physical environment maps based on terahertz communication signals faces several significant challenges, which can be summarized in the following three aspects:

[0005] First, constructing radio maps presents challenges. On one hand, acquiring a radio map of a specific scene typically requires deploying receivers across the entire scene. However, due to hardware costs and real-world environmental conditions, this approach is not feasible. Therefore, radio map construction can only rely on data acquired from a limited number of receivers, facing the challenge of data sparsity. On the other hand, terahertz signals have extremely narrow beamwidths. To generate a radio map of the entire scene, an omnidirectional beam scan of the scene is required. However, THz ISAC technology has stringent requirements for response time, and the omnidirectional scan and subsequent data processing are too time-consuming, failing to meet the real-time requirements of ISAC.

[0006] Second, the construction of physical environment maps also faces multiple constraints. First, the high path loss, short transmission range, and extreme channel sparsity of terahertz frequencies compared to sub-6 GHz bands limit the ability to accurately detect and interpret environmental features over large areas, resulting in highly sparse and unevenly distributed sensor data. This data often fails to provide a global representation of the environment, making the task more difficult. Second, the increasing complexity of modern environments—such as dense urban scenes containing numerous obstructions, scatterers, and interference sources—further increases the complexity of physical environment maps.

[0007] Third, there is a lack of effective solutions for the joint construction of radio maps and physical environment maps. Traditional model-driven methods require pre-setting propagation parameters, but the materials of obstacles and signal propagation mechanisms in the real environment are complex and variable, making it difficult for pre-set parameters to match the real scene, leading to errors in model estimation. Furthermore, given the numerous difficulties in independently constructing radio maps and physical environment maps, the difficulty of jointly constructing them is further increased.

[0008] Traditional radio mapping solutions can be categorized into three types based on their technical approaches: model-based methods, data-based methods, and a combination of model and data-driven approaches. However, none of these three approaches are well-suited to the needs of terahertz communication scenarios. Model-based methods, represented by statistical channel models and deterministic ray tracing models, cannot accurately depict scene-specific characteristics such as strong obstruction from concrete walls and strong absorption at specific frequencies by air molecules. They exhibit significant perception errors in areas with dense obstacles. While deterministic ray tracing models can describe propagation details, they require high-precision geometric data of the environment, resulting in a significant increase in computational complexity. They become completely ineffective in unknown scenarios or with sparse data. Data-based methods rely on densely deployed sensors and omnidirectional scanning, with hardware and power costs far exceeding the requirements for high efficiency. Early model-and-data-driven approaches were often simple combinations of physical models and shallow machine learning, such as using ray tracing simulation data to train networks or empirical formulas to correct outputs. However, the differences between simulation and real-world scenarios, and the inability of empirical formulas to reflect the essential laws of electromagnetic wave propagation, lead to misjudgments.

[0009] Current technical solutions for physical environment map construction are relatively limited, mainly falling into three categories: target localization and tracking-oriented methods, Simultaneous Localization and Mapping (SLAM) technology, and scattering point and camera image fusion methods. None of these methods are suitable for the application requirements of terahertz communication scenarios. Target localization and tracking-oriented methods focus on the localization and tracking of targets within the environment but do not incorporate the construction of a complete physical environment map, thus failing to support the global environmental semantic representation required for terahertz communication. SLAM technology relies on Channel State Information (CSI) parsing or signal processing to extract environmental scattering points, thereby achieving environmental modeling. However, this technology has limitations; it can either only generate two-dimensional environmental images or only perceive a few scattering points, failing to accurately reconstruct the environment. While scattering point and camera image fusion methods can improve environmental reconstruction to some extent, their applicability is limited, only supporting obstacle modeling for a few specific materials and surface types. When faced with diverse object shapes and material differences in complex scenes, they cannot achieve universal physical environment map construction.

[0010] Current methods for jointly constructing radio maps and physical environment maps largely rely on deep learning, but they generally lack prior physical knowledge and depend solely on data-driven modeling. This presents challenges in terahertz communication scenarios: First, receiver data exhibits severe sparsity. Terahertz base stations have narrow beamwidths, requiring the use of a large number of beams... It takes several beams to cover space, The beamwidth is limited, but due to the cost of base station transmit power and the time cost of omnidirectional scanning (real-time sensing requirements), the base station can only choose the total number of sparse beams. Beam scanning is performed in key directions. The scaling factor ranges from 0 to 1; simultaneously, the limited number of receivers in the scene further exacerbates the data sparsity. Secondly, the mapping complexity is high. The nonlinear propagation characteristics of terahertz signals, such as molecular absorption and obstacle reflection, along with the limited power of base stations leading to significant noise impact, further increase the mapping complexity from the received signal strength RSS value to the radio map and the physical environment map. Finally, the first two factors interact significantly: data sparsity makes it difficult for the model to capture key signal features under nonlinear propagation characteristics, while the complex mapping further amplifies the impact of data sparsity and noise, causing the model to fail to accurately learn the relationship between the signal and the environment. These combined problems lead to insufficient accuracy in data-driven modeling in terahertz scenarios. To address this, a joint construction method incorporating physical propagation laws is proposed. The physical mechanism of terahertz signals is embedded as a constraint in the model, achieving synergy between data-driven approaches and physical priors, ensuring high-precision construction of both types of maps.

[0011] From a physical perspective, terahertz signals, as a type of electromagnetic wave, strictly follow the Helmholtz equations derived from Maxwell's equations in their propagation characteristics. The frequency domain form of these equations is... ,in For wave number, Angular frequency, At the speed of light, For the Laplace operator, The imaginary unit, Permeability, The electric field intensity is a vector field. The current density can accurately characterize the amplitude and phase changes of the signal, as well as the reflection and refraction at obstacle boundaries. In two-dimensional terahertz communication scenarios, for transverse electromagnetic waves (TE mode), the Helmholtz equations can be simplified to a scalar form. , The electric field strength is a scalar field. After being discretized by the central difference method, it can be directly used to quantize the distribution of signal strength in the grid space.

[0012] From the perspective of the essence of the problem, the construction of terahertz dual maps is a typical generative problem rather than a discriminative problem: the omnidirectional RSS to be generated is distributed in the obstacle layout and is not included in the sparse input data. The model needs to actively complete the global features based on local information, and the output is a continuous value with multimodal possibilities. During training, the mask content also needs to be recovered based on sparse data (which can be regarded as mask samples of the complete radio map). This is highly consistent with the core logic of generative models. Therefore, generative models have become the inevitable choice to solve this problem.

[0013] In exploring the applications of generative models, flow matching models have gradually gained attention due to their unique advantages. Compared with traditional generative models, flow matching models learn a continuous and reversible transformation flow from a simple prior distribution to the target RSS distribution, avoiding mode collapse without adversarial training, and their inference efficiency is improved compared to ordinary diffusion models, demonstrating potential for adaptation to terahertz scenarios. However, the application of generative models still faces technical bottlenecks: on the one hand, training is significantly difficult, as the sparsity of RSS data and the nonlinear propagation characteristics of signals in terahertz communication scenarios make it difficult for models to learn stable global mapping rules; on the other hand, the integration of physical mechanisms is difficult, as existing models mostly rely on pure data-driven learning of the statistical characteristics of the RSS distribution, failing to deeply integrate the physical laws of terahertz signal propagation into the generation process, leading to physical inconsistencies in the generation results. Summary of the Invention

[0014] The technical problem to be solved in this application is to effectively embed the physical mechanism of terahertz signal propagation into the model framework while reducing the training difficulty of generative models, so as to meet the needs of obstacle perception and radio map reconstruction. A solution is proposed to generate high-precision omnidirectional radio maps and obstacle layout by integrating Helmholtz equation constraints, regionalized power boundary constraints and classifier-free flow matching model.

[0015] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for constructing a terahertz environment semantic map based on a physical information flow matching model, comprising the following steps:

[0016] Training dataset construction steps: Obtain a sparse directional beam received signal strength dataset collected from real-world scenarios. Real omnidirectional radio map and real physical environment Construct training set ;

[0017] Flow matching model construction and training steps: Construct a flow matching model with a built-in dynamic determination module for three types of propagation regions. This model uses the sparse directional beam received signal strength dataset. As input conditions, the actual omnidirectional radio map is used. and real physical environment To supervise the target and incorporate dual physical constraints during training, the optimal model parameters are obtained. ;

[0018] Omnidirectional radio map and physical environment generation steps: Acquire sparse directional beam received signal strength data in real-world application scenarios. The data is then input into a pre-trained flow matching model, which generates omnidirectional radio signals for the current scene through classifier-free inference. With physical environment .

[0019] This invention addresses the core challenges of obstacle perception in terahertz communication scenarios—the difficulty in generating high-precision omnidirectional radio maps and physical environments from sparse directional beam RSS data, the lack of physical consistency in generative model outputs, and high hardware and time costs—by proposing a flow matching model. The overall solution is based on the core logic of "real data-driven + dual physical constraints + built-in region determination," achieving its goal through three core steps: First, a training set of "directional beam RSS data - omnidirectional radio map - obstacle layout map" collected from real-world scenarios is constructed. Second, a flow matching model with built-in dynamic determination modules for three types of propagation regions is designed, incorporating dual physical constraints and undergoing training. Finally, high-precision omnidirectional radio maps and obstacle layouts are generated based on sparse test data, balancing generation accuracy, physical consistency, and deployment costs.

[0020] The beneficial effects of this invention are as follows:

[0021] 1. Deep integration of region determination and model: It has three built-in dynamic region determination modules, which do not require additional preprocessing. It completes region division in real time based on beam direction, coordinates and RSS features, providing accurate basis for regional power constraints and improving physical consistency.

[0022] 2. Dual constraints and guidance work synergistically: Helmholtz equation constraints ensure compliance with macroscopic laws of signal propagation, regional power constraints refine microscopic characteristics based on dynamic judgment results, classifier-free guidance enhances data fitting accuracy, and combined with real-scene supervision, the generation achieves both high accuracy and physical rationality.

[0023] 3. Data and cost advantages: Directly uses real-world collected data, eliminating the need for ray tracing; only requires... Dual outputs and region determination can be achieved using only one directional beam data;

[0024] 4. High generalizability and practicality: The regional determination logic is adaptable to different scenarios, and the dual constraints reduce dependence on specific data. It can be directly applied to complex terahertz communication scenarios such as urban commercial areas and industrial parks. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the flow matching model based on dual physical constraints and classifier-free guidance of the present invention, which sequentially shows the four core steps of training dataset construction, model training, model inference, radio graph and obstacle layout generation;

[0026] Figure 2 This is a top-down view of a scenario according to an embodiment of the present invention. 1 represents a high-rise office building (3 buildings), 2 represents a low-rise shop (12 shops), and 3 represents a terahertz base station (located in the center of the scenario). The angle between the beam direction and the line connecting the base station and the grid is also marked.

[0027] Figure 3 This is a schematic diagram of the training process of the present invention. Pure noise is used as the initial input. Under the guidance of sparse beam RSS, obstacle distribution map and radio map are generated by a flow matching model that incorporates region correction and Helmholtz equation. Then, the generated image is compared with the real obstacle distribution map and radio map to calculate the error, which is used as the parameter for the optimization model feedback.

[0028] Figure 4 This is a schematic diagram of the reasoning process of the present invention. Pure noise is input into the trained flow matching model, and under the guidance of sparse beam RSS, an obstacle distribution map and a radio map are generated.

[0029] Figure 5 This is a schematic diagram of the three-area determination process of the present invention, which sequentially shows the three-step logic of potential line-of-sight screening, occlusion detection, and RSS threshold division. Detailed Implementation

[0030] The method for constructing terahertz environment semantic maps includes: 1. Training dataset construction steps, 2. Flow matching model construction and training steps, and 3. Omnidirectional radio map and physical environment generation steps.

[0031] I. Steps for constructing the training dataset:

[0032] The training set is constructed by directly using three types of core data collected from real-world scenarios, with each scenario corresponding to a specific one:

[0033] 1a. Acquisition and Rasterization of Realistic Obstacle Layout: Select a typical terahertz communication scenario. First, rasterize the scenario according to a preset precision, dividing the scenario into a uniform grid (e.g., dividing the scene into a grid of uniform obstacles). The scene according to The precision is rasterized, and it can be divided into: (each grid has a unique index and center coordinates). , The grid index is then used; subsequently, high-precision mapping is employed to determine the location, geometry (to determine the grid area they occupy), and relative permittivity of obstacles within the scene. To create a real physical environment ;when Represents grid The area is an obstacle zone. Represents grid This is a free space region;

[0034] 1b. Acquisition of Real Radio Map: A terahertz base station is deployed in the center of the scene. The base station performs omnidirectional beam scanning and collects the RSS values ​​of all grids through receivers densely distributed throughout the scene to generate an omnidirectional real radio map. ;

[0035] 1c. Sparse directional beam RSS data acquisition: In the same scenario, the same base station is selected. A set of preset directional beam directions, wherein the directional beam direction is composed of the corresponding beam center direction angle set. This indicates that the base station is configured according to the direction of each beam. To propagate signals, Total number of sparse beams, beam direction index Within the scene, receivers measure RSS in each directional beam direction. Due to a limited number of receivers, the acquired RSS data is sparsely distributed in space. Define the first... The receiver is at position coordinates No. The RSS value measured in each beam direction is ,index , If the total number of receiver nodes deployed in the scenario is [the number of nodes deployed in the scenario], then in the [number of nodes deployed in the scenario]... The RSS superposition value of sparse directional beams of a receiver is defined as the energy superposition of signals in each beam direction. , represented as:

[0036] ;

[0037] This ultimately forms a sparse directional beam RSS dataset. ;

[0038] 1d. Training set construction: using real physical environments Real radio maps and Sparse Directional Beam RSS Dataset The training set is constructed by matching each scenario one by one. , The total number of scenes, For scene indexing, For the first Sparse directional beam RSS dataset for various scenarios For the first The real physical environment of each scene. For the first A real radio diagram of a scene.

[0039] II. Steps for building and training a flow matching model:

[0040] 2a) Core definition of the flow matching model: Based on the flow matching framework, the time is defined. Stream variables during time , , where is a continuous-time variable; hour, It is pure noise, follows a standard Gaussian distribution, and has dimensions consistent with real radio maps and real physical layouts; hour, ;

[0041] The flow matching model employs a dual-output U-Net architecture and aims to use the sparse directional beam RSS dataset. Given input conditions, learn to generate guided algorithms from pure noise. Mapping to target continuous flow vector field A vector field is defined as a continuous flow. Regarding time The derivative, This represents the model parameters of the flow matching model; the flow matching model also outputs a generated radio graph. With the generation of physical environment It also has a built-in module for determining three types of propagation regions; the classification-free guide mechanism improves the accuracy of model mapping through reasoning guidance.

[0042] 2b) Forward Flow Definition: Constructing a resolvable forward flow Describes the continuous transition from pure noise to the target output:

[0043] ;

[0044] in, hour , It is pure noise. hour , indicating the target output; During the incremental process, the noise component gradually decreases while the true RSS feature gradually increases, ensuring strict alignment between the forward flow and the supervised target; its vector field can be expressed as:

[0045] ;

[0046] 2c) Dynamic determination of three types of propagation regions: During model training and inference, three types of regions are determined in real time based on beam direction, coordinate relationship, and RSS value features. The specific steps are as follows:

[0047] (1) Targeting A directional beam with different directions, the center direction angle of which is The beamwidth is Calculate the angle between the base station-grid connection line and the positive direction of the horizontal axis. subscript Indicates base station, subscript Represents a grid. For base station coordinates, For raster coordinates; if Then the grid Potential line-of-sight area If yes, proceed to step (2); otherwise, proceed to step (3).

[0048] (2) Obtain the set of all grids through which the connection between the base station and the grid passes. ,like The grid without obstacles, i.e., all ,and satisfy Then determine the grid For viewing distance area Otherwise proceed to step (3); For free space receiving power, This is the multipath enhancement factor. For located in the grid The received signal strength RSS measured by the receiver at the location;

[0049] (3) If , The minimum power required for complete blockage, and which is satisfied in all beam directions, is considered a blocked area. ;like And there is a signal only in some beam directions. If the attenuation coefficient is non-line-of-sight, then the grid is determined. Non-line-of-sight area ;

[0050] 2d) Incorporating dual physical constraints:

[0051] (1) Helmholtz equation constraint: Based on the TE mode of terahertz signal, the Helmholtz equation is simplified into scalar form. Discretize it into a grid using the central difference method. residual at the location :

[0052] ;

[0053] in, Generate a raster for the flow matching model The RSS value at that location For scene grid spacing, For grid Wavenumber squared; where, free space region Obstacle area , Let be the relative permittivity of the obstacle. For free space wavenumber;

[0054] (2) Regionalized power boundary constraints: Based on the three types of regions determined in step 2c), differentiated power boundaries are set, and the RSS value exceeding the boundary is penalized by a loss function: The boundary is , The boundary is , The boundary is , To stabilize training of small positive numbers;

[0055] 2e) Classifier-free guided training mechanism:

[0056] The core of classifier-free guided training lies in achieving strong guidance of the generation process by jointly training both conditional and unconditional generative models. During the training phase, the sparse directional beam RSS dataset is used. As input for training conditions, the model learns both conditional RSS distribution fitting and unconditional distribution fitting; during the generation phase, the difference between the outputs of the two types of models is used to guide the sampling process, thereby more accurately approximating the real radio graph distribution under sparse beam constraints.

[0057] 2f) Construction of the total loss function: The total loss function is constructed by combining the weights of multiple loss terms, balancing the flow matching direction, physical constraints, data supervision, three-class region discrimination, and the need for classifier-free guidance. :

[0058] ;

[0059] , , , These are the weighting coefficients for the Helmholtz equation residual loss, boundary constraint loss, supervision loss, and regionalized power loss, respectively.

[0060] It is the flow matching loss, used to minimize the vector field generated by the flow matching model. Deviation from the reference vector field:

[0061] ;

[0062] in, Expressing expectations, It is pure noise. The forward flow derivative serves as the directional reference. Indicates from the training set A sparse beam data sampled in the middle is Real radio diagrams are And the real physical environment is ; This represents a continuous-time variable randomly sampled from a uniform distribution over the interval [0,1]. Represents the square of the L2 norm;

[0063] The residual loss from the Helmholtz equations is used to ensure that the generated RSS value conforms to the laws of electromagnetic wave propagation.

[0064] ;

[0065] Where H and W represent the total number of grid cells in the scene grid in the length and width directions, respectively;

[0066] It is a boundary constraint loss, used to ensure that the RSS value at the obstacle boundary conforms to physical laws:

[0067] ;

[0068] in, It is a set of obstacle boundary meshes. A grid located at the boundary of an obstacle. The theoretical RSS value that should be met;

[0069] It is the supervised loss, used to measure the difference between the model output and the supervised objective:

[0070] ;

[0071] in, The radio graph generated for the flow matching model, The physical environment generated for the flow matching model. The square of the Frobenius norm;

[0072] To regionalize power loss, and to ensure that the RSS values ​​of different regions are correctly corrected:

[0073] ;

[0074] in, Indicates taking the positive part. It is an L1 norm. and These are the upper and lower boundary matrices of the region power, respectively;

[0075] 2h) Flow matching model training: The flow matching model adopts a dual-output U-Net architecture. The input layer is a concatenated feature set of "pure noise + sparse directional beam RSS dataset". The intermediate layer enhances feature extraction through residual connections and attention mechanisms. The output layer simultaneously outputs an omnidirectional radio graph. With physical environment Finally, the optimal model parameters are obtained. The stream matching training process is as follows:

[0076] (1) Constructing training samples: Randomly select a sample from the training set T. It contains three parts of sparse directional beam RSS data. (As a conditional input) Real omnidirectional radio map And the real physical environment (obstacle layout) ;

[0077] (2) Random sampling time point: sampling at a random time point , =0 represents a pure noise state. =1 represents the true state of the data, and the intermediate value represents the mixed state of noise and data;

[0078] (3) Construct the current state Sample pure noise from a standard Gaussian distribution. ,according to The noise and real data are linearly mixed to obtain the current time. status That is, the flow variable at time 1 The value of ; this state is the noisy input that the flow matching model needs to process at the current time step; , With a mean of 0 and a variance of The standard multivariate normal distribution;

[0079] (4) Prediction step: The noisy input With conditional information A U-Net neural network with two inputs and two outputs is used; the model outputs two results: the predicted vector field. (i.e., "which direction should we go next"); it also outputs the current predicted radio chart. and physical environment (Used to calculate monitoring loss);

[0080] (5) Calculate the total loss:

[0081] The total loss is obtained by weighted summation of multiple parts:

[0082] ;

[0083] (6) Backpropagation and parameter update: Calculate the gradient based on the total loss. Update neural network parameters using gradient descent. Repeat the above steps until the model converges, and output the optimal model parameters. .

[0084] III. Steps for generating omnidirectional radio maps and physical environments:

[0085] 3a) Directional beam RSS acquisition during the online application phase:

[0086] After the model completes offline training and obtains optimal parameters, it is deployed to a terahertz base station as the base station's radio environment perception module. During actual operation after deployment, no additional training or parameter updates are performed on the model. Instead, data collected by the base station in real communication scenarios is directly used as the model input to achieve perception of the radio environment in the scenario.

[0087] Specifically, in the terahertz communication scenario to be perceived, the base station follows... Signal transmission and reception are performed using a preset directional beam direction. Receivers sparsely distributed within the scene measure the RSS and transmit the measured data back to the base station. The base station processes the received RSS data and calculates the RSS superposition value for each receiver node, thus forming the sparse directional beam RSS data for the current scene. The obtained data is then constructed using the method described earlier for constructing a sparse directional beam RSS dataset, forming an application dataset. This dataset serves as a conditional input to a deployed flow matching model, driving the model to infer and generate the radio map and obstacle layout for the current scene.

[0088] 3b) Classifier-free guided inference: treating pure noise and training conditions Input optimal model ,control The value is incremented gradually from 0 to 1; to conceptually distinguish between the construction path during training and the generation path during inference, a superscript ~ is added to the state during inference, and the stream variable is updated using the Euler method:

[0089] ;

[0090] when At the same time, the model outputs an omnidirectional radio graph of the real application scenario. With physical environment .

[0091] The process of generating stream matching inference without classifier guidance is as follows:

[0092] (1) Receive inference dataset Euler steps : This refers to sparse directional beam RSS data collected in real-world application scenarios; the Euler step count is used to summarize the entire generation process from... arrive The number of steps in the division;

[0093] (2) Reasoning steps:

[0094] Initialization: Set the current time ; Calculate step size (unit time zone) Divide evenly (Small step); Randomly sample a pure noise sample from a standard normal distribution. , as the starting point of the generation process;

[0095] Step-by-step iteration: for the count variable Repeat the following operations:

[0096] Predicting direction: the current state and conditional input Both are fed into the trained flow matching model; the model outputs the vector field at the current time step. This vector field is used to tell which direction to adjust;

[0097] Update status: Take a small step forward according to Euler's method: This is equivalent to updating the current state to the state of the next time step, following the direction indicated by the model.

[0098] Forward Time: Update current time: ;

[0099] Complete generation: When the loop ends At that time, we obtained the final state. This final state contains two parts of information: the inferred omnidirectional radio graph. and physical environment .

[0100] To enable those skilled in the art to better understand the present application, the following will refer to the accompanying drawings in the embodiments of the present application (…). Figures 1-5 This application provides a clear and complete description of the technical solutions in its embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0101] It should be noted that the core of the flow matching model proposed in this application, which integrates dual physical constraints and classifier-free guidance, is through... Coupled with the physical propagation laws of sparse beam RSS data in each direction and terahertz signals, and combined with the division of three types of regions, accurate generation of scene radio maps and obstacle layouts is achieved; the following examples use... Figure 2shown Taking an urban commercial area scenario as an example, the dataset is constructed directly using real radio maps of the cumulative RSS values ​​of four directional beams, real omnidirectional radio maps, and real obstacle layout maps. The model has a built-in area determination module to achieve synchronous output of radio maps and obstacle layouts. The following embodiments, together with the accompanying drawings, illustrate the complete implementation process of the technical solution.

[0102] Example

[0103] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the present application. The method specifically includes the following steps:

[0104] Step S1: Training Dataset Construction

[0105] This step aims to obtain one-to-one corresponding samples of the real physical environment, radio graph, and RSS superposition values ​​of the four directional beams, providing supervision and physical constraints for model training. Specifically, it includes sub-steps S1.1 to S1.4.

[0106] S1.1 Acquisition of Real Physical Environment

[0107] 1. Scene rasterization: [This likely refers to a specific process or feature, but the full context is unclear without further information.] Figure 2 As shown Urban commercial district scene Precision rasterization, to obtain Number of grids, total number of grids The center coordinates of each grid cell are marked as follows: ( , , (These are planar coordinates, in meters). For raster indexes;

[0108] 2. Obstacle mapping: Obstacle information is clearly defined through radar mapping, especially for high-rise office buildings ( Figure 2 Marker 1) indicates three rectangular buildings located in the northeast, northwest, and south of the scene. They are made of concrete and have a relative permittivity of [missing information]. The area occupied by the grid is marked as Low-rise shops ( Figure 2 (2) There are 12 rooms in total, distributed along two main east-west and north-south roads. The material is a mixture of brick and stone, and the relative permittivity is... The area occupied by the grid is marked as ;

[0109] 3. Realistic Environment Generation: Integrating the grid positions and material parameters of the obstacles mentioned above to form a realistic physical environment. ,in, Represents grid The area is an obstacle zone. Represents grid This is a free space region.

[0110] S1.2 Acquisition of Real Radio Maps

[0111] 1. Base station and signal parameter settings: Terahertz base stations are deployed in Figure 2 The four grids in the center of the scene marked with 3 carrier frequency Transmission power Omnidirectional beam scanning is performed, with the receiver located at the center of each grid.

[0112] 2. Receiver Data Acquisition: 40,401 receivers are densely deployed to collect the RSS value of each grid cell, generating a radio map. .

[0113] S1.3 Directional Beam RSS Superposition Value Acquisition

[0114] In real-world scenarios, receiver distribution is often random and limited by various factors, resulting in a finite number of receivers. Therefore, to enable the trained model to adapt to the wireless environment perception requirements under different receiver distributions, it is necessary to simulate the sparse observation conditions formed by the limited number of receivers deployed in real-world scenarios at the data level.

[0115] Therefore, during the scene construction phase, RSS data covering all grid locations is first acquired by densely deploying receivers. Based on this, to reduce beam scanning time and transmit power overhead, a scaling factor k of 0.22 is selected, corresponding to selecting four directional beam directions from the omnidirectional beam set.

[0116] exist Figure 2 In the same scenario and at the same base station location, four beam directions were selected (the beam center angles are respectively...). , , , The base station transmits signals according to the directional beam direction, and each receiver in the scene collects RSS data and transmits it back, calculating the RSS superposition value of each receiver under the directional beams in the four directions.

[0117] After the RSS superposition value is calculated, not all the receiver data is directly used for model training. Instead, a portion of the receiver samples are selected from the receivers according to a preset ratio to achieve sparsification and form a sparse directional beam RSS dataset for input model.

[0118] S1.4 Training Set Assembly

[0119] Will Corresponding to each scenario, expand to 400 scenarios with different obstacle layouts to form a training set

[0120] Step S2: Construction and training of the flow matching model

[0121] This step is the core technical link. Combining Figure 3 the training process shown, through defining flow variables, building three types of region dynamic determination modules, integrating dual physical constraints, constructing multi-loss functions, and training without a classifier guidance, precise generation of radio maps and obstacle layouts is achieved. The specific steps include sub-steps S2.1 to S2.5.

[0122] S2.1 Model core definition

[0123] Based on the flow matching framework, clarify the input and output logic of the model:

[0124] 1. Definition of flow variable: The flow variable , when (pure Gaussian noise), when ;

[0125] 2. Model objective: Learn the vector field conditioned on , and output the radio map and the obstacle layout .

[0126] S2.2 Forward flow definition

[0127] Construct the forward flow , ensuring that the flow transformation smoothly transitions from noise to the target scenario.

[0128] S2.3 Build three types of region dynamic determination modules

[0129] As Figure 5 shown, the following determinations are performed during model training:

[0130] 1. Potential line-of-sight screening: Calculate (base station coordinates ), if , classify it as ;

[0131] 2. Occlusion detection and determination: Obtain the set of connected grid . If there is no obstacle grid and satisfies , determine it as ;

[0132] 3. and Division: And all beams satisfy the condition, thus it is determined to be ; Furthermore, some beams have signals, which is determined to be... .

[0133] S2.4 Dual Physical Constraints Integration

[0134] Helmholtz equation constraints:

[0135] 1. Equation simplification: The Helmholtz equation in the TE mode is as follows: , ;

[0136] 2. Wavenumber Correction: Free Space Obstacle area (Since transmission is not considered, the obstacle region except for the obstacle boundary can be regarded as an ideal conductor, the field distribution is 0 everywhere, the wave number loses its meaning, and the field strength can be directly regarded as 0).

[0137] 3. Residual formula:

[0138] ;

[0139] .

[0140] S2.4 Regionalized Power Constraint

[0141] Set boundaries based on the judgment results. boundary , boundary , boundary .

[0142] S2.6 Loss Function and Model Training

[0143] ,in, It is a flow matching loss, which ensures that the learned flow is consistent with the actual direction; It is the Helmholtz residual loss, and the result of forced generation conforms to physical laws; It is a boundary constraint loss, ensuring that the RSS at the obstacle boundary is reasonable; It is supervised loss that makes the generated results fit the true labels better; To address regionalized power loss, ensure that the RSS values ​​of different regions are correctly corrected;

[0144] use Figure 3The training process shown is a single-machine centralized training. The input layer concatenates pure noise and RSS data from four directions. The pure noise serves as the starting data for the model to generate obstacle layouts and radio graphs, while the RSS data from the four directions... As a guiding condition for the training process, the intermediate layer includes residual connections, attention mechanisms, and region judgment modules to enhance feature extraction, and the output layer 1 is... Output layer 2 is .

[0145] Step S3: Generating Radio Map and Obstacle Layout

[0146] S3.1 Inference Data Acquisition

[0147] In a real-world scenario, the superimposed RSS data values ​​of a number of receivers under the four directional beam directions are collected to obtain... .

[0148] S3.2 Model Reasoning

[0149] Will and Input model, t increments from 0 to 1, output and ,like Figure 4 .

[0150] The method described in this embodiment has the following effects:

[0151] 1. Increased generation accuracy: Incorporating dual physical constraints, the RSS value is forced to follow the laws of electromagnetic wave propagation to avoid physical inaccuracies; the dual-output U-Net coupled beam and RSS features accurately distinguish between obstacles and shadow areas, resulting in more accurate reconstruction of building layouts.

[0152] 2. Improved inference efficiency: The flow matching model does not require multiple iterations. The dual-output architecture simultaneously produces an omnidirectional graph and obstacle layout, saving additional subsequent processing and adapting to real-time perception.

[0153] 3. Reduced cost and energy consumption: Only sparse RSS data in 4 directions is required, eliminating the need for omnidirectional beam scanning and dense receiver deployment, reducing hardware investment and base station power consumption, and adapting to existing infrastructure.

[0154] 4. Enhanced generalization ability: Physical constraints are universal, reducing dependence on specific scenario data. When faced with changes in obstacle layout, there is no need for large-scale retraining, adapting to different scenarios.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0156] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a terahertz environment semantic map based on a physical information flow matching model, characterized in that, Includes the following steps: Training dataset construction steps: Obtain a sparse directional beam received signal strength dataset collected from real-world scenarios. Real omnidirectional radio map and real physical environment Construct training set ; Flow matching model construction and training steps: Construct a flow matching model with a built-in dynamic determination module for three types of propagation regions. This model uses the sparse directional beam received signal strength dataset. As input conditions, the actual omnidirectional radio map is used. and real physical environment To supervise the target and incorporate dual physical constraints during training, the optimal model parameters are obtained. ; Omnidirectional radio map and physical environment generation steps: Acquire sparse directional beam received signal strength data in real-world application scenarios. The data is then input into a pre-trained flow matching model, which generates omnidirectional radio signals for the current scene through conditional generation inference based on classifier-free guided sampling. With physical environment .

2. The method as described in claim 1, characterized in that, The training dataset construction steps further include: Realistic physical environment acquisition: The scene is rasterized to obtain the coordinates of each grid cell. The location, geometry, and relative permittivity of obstacles were determined through mapping. To create a real physical environment ,when Represents grid The area is an obstacle zone. Represents grid This is a free space region. For grid indexing; Realistic omnidirectional radio map acquisition: Deploy terahertz base stations to perform omnidirectional beam scanning, and collect the received signal strength values ​​of all grids through densely distributed receivers to generate a realistic omnidirectional radio map. ; Sparse directional beam received signal strength data acquisition: The base station selects n preset directional beam directions for signal propagation, and collects the received signal strength values ​​in each direction using sparsely distributed receivers, calculating the energy superposition value of each receiver. To form a sparse directional beam received signal strength dataset ;in, Indicates the first in the target scene The receiver is at position coordinates Place, No. The RSS values ​​measured in each beam direction, index , This is the total number of receiver nodes deployed in the scenario; Training set construction: This involves constructing the real physical environment within the same scene. Real omnidirectional radio map Data set of sparse directional beam received signal strength One-to-one correspondence, forming the training set ,in, The total number of scenes, For scene indexing, For the first Sparse directional beam RSS dataset for various scenarios For the first The real physical environment of each scene. For the first A real radio diagram of a scene.

3. The method as described in claim 1, characterized in that, The steps for building and training the flow matching model further include: Core definition of the flow matching model: Definition time Stream variables during time ,in , It is pure noise. Construct a dataset of received signal strength using sparse directional beams. Vector field with conditions , This represents the model parameters of the flow matching model. Indicates time The streaming variable state is obtained by using a dual-output U-Net architecture to synchronously output the generated radio graph. With physical environment ; Forward flow definition: Define a resolvable forward flow. Its vector field is ; Dynamic determination of the three propagation areas: During model training and inference, the grid is determined as the line-of-sight area in real time based on the beam direction, base station-grid geometry, and received signal strength. Non-line-of-sight area or obstructed area ; Incorporating dual physical constraints: including Helmholtz equation constraints used to constrain the received signal strength values ​​generated by the flow matching model and regionalized power boundary constraints used to constrain the received signal strength values ​​generated by the flow matching model within the corresponding propagation region; Among them, the constraints of the Helmholtz equation will simplify the scalar form of the Helmholtz equation. Discretize into residuals using the central difference method. Regionalized power boundary constraints are set based on three types of regions determined dynamically, with differentiated power boundaries. Constructing the total loss function: using flow matching loss Helmholtz equation residual loss Boundary constraint loss , monitoring losses and regionalized power loss Constructing total loss : ; in, , , , These are the weighting coefficients for the Helmholtz equation residual loss, boundary constraint loss, supervision loss, and regionalized power loss, respectively. Model training: Gradient descent is used to optimize the total loss function and update the model parameters. until convergence is achieved and the optimal model parameters are obtained. .

4. The method as described in claim 3, characterized in that, Residuals in Helmholtz Equation Constraints for: ； in, Generate a raster for the flow matching model The RSS value at that location For scene grid spacing, For grid Wavenumber squared; where, free space region Obstacle area , Let be the relative permittivity of the obstacle. is the free space wavenumber.

5. The method as described in claim 3, characterized in that, Stream matching loss for: ； in, Expressing expectations, The forward flow derivative serves as the directional reference. Indicates from the training set A sparse beam data sampled in the middle is Real radio diagrams are And the real physical environment is ; This represents a continuous-time variable randomly sampled from a uniform distribution over the interval [0,1]. This represents the square of the L2 norm.

6. The method as described in claim 3, characterized in that, Helmholtz equation residual loss for: ； Where H and W represent the total number of grid cells in the scene grid in the length and width directions, respectively.

7. The method as described in claim 3, characterized in that, Boundary constraint loss for: ； in, It is a set of obstacle boundary meshes. A grid located at the boundary of an obstacle. The theoretical RSS value that should be met.

8. The method as described in claim 3, characterized in that, Monitoring losses for: ； in, It is the square of the Frobenius norm.

9. The method as described in claim 3, characterized in that, Regionalized power loss for: ； in, Indicates taking the positive part. It is an L1 norm. and These are the upper and lower boundary matrices of the region power, respectively.

10. The method as described in claim 1, characterized in that, The determination logic of the three-type propagation region dynamic determination module includes: Potential line-of-sight screening: Calculate the angle between the base station-grid connection and the positive direction of the horizontal axis. , For grid indexing, For beam direction index, if the corner is in any preset beam direction beamwidth If it is within the range, it is marked as a potential line-of-sight candidate; Occlusion detection and line-of-sight determination: Obtain the set of all grid cells traversed by the line connecting the base station and the grid. ,like A grid without obstacles in the middle and the received signal strength of the grid satisfies the free space power constraint. Then it is determined to be the line-of-sight region. ;in, For located in the grid The received signal strength RSS measured by the receiver at the location; Non-line-of-sight and obstruction determination: If the received signal strength of the grid is lower than the minimum power for complete obstruction. If there is no signal in any beam direction, it is determined to be an obstructed area. If the received signal strength value meets the following conditions: If there is a signal only in some beam directions, it is determined to be a non-line-of-sight area. .

Images