Indoor radio map construction method and device based on physical enhancement diffusion model

Abstract

The invention discloses an indoor radio map construction method and device based on a physical enhancement diffusion model, and the method comprises the steps: constructing initial discontinuity physical priori including a diffraction point set and a transmission boundary point set according to the geometric information and building material information of an indoor environment; pruning the diffraction point set in the initial discontinuous physical prior to update the discontinuous physical prior; generating a discontinuity physical prior graph according to the indoor radio access point and the updated discontinuity physical prior; constructing a multi-modal physical priori including a coordinate position of an indoor radio access point, a reflection coefficient graph and a transmission coefficient graph representing a building material at each position, and a discontinuous physical priori graph; and processing the multi-modal physical priori by a multi-modal fusion encoder, and inputting a processing result as a model input condition into the trained decoupling diffusion model to generate an indoor radio map. According to the invention, on-site measurement is not needed, and the high-fidelity indoor radio map can be quickly generated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication, and particularly relates to a method and device for constructing an indoor radio map based on a physical enhanced diffusion model. BACKGROUND

[0002] Indoor positioning plays a key role in intelligent buildings, industrial automation, emergency response, and immersive applications, and there is an increasing demand for millimeter to sub-millimeter level positioning accuracy and scalable, low-latency deployment. Radio Map (RM) as an electromagnetic representation of environment perception can associate scene geometry, material properties and signal strength spatial distribution, thus enabling high-precision positioning without expensive on-site surveying, and is a key technology for the next generation of wireless communication and positioning services.

[0003] Currently, the technical solutions for constructing high-precision indoor radio maps mainly fall into three categories: Distance measurement model-based methods: This type of method relies on received signal strength information (RSSI) or time of arrival (ToA) measurements and combines statistical distance loss models to estimate distance. For example, by using the free space path loss model or the logarithmic distance path loss model, the received signal strength is converted into the distance from the source of transmission, thereby enabling positioning.

[0004] Fingerprint library-based methods: This method requires extensive offline surveying, densely collecting a large number of signal fingerprints (such as RSSI, CSI, etc.) at location points in the target area, and constructing a "location-fingerprint" database. In positioning, the real-time measured signal fingerprint is matched with the records in the database (such as using the K-nearest neighbor algorithm) to determine the user's location.

[0005] Simulation-based radio map construction methods: This method aims to avoid large-scale field measurements and generates radio maps through software simulation. Specifically, it can be divided into: 1) Full-wave electromagnetic solver: This method accurately simulates electromagnetic wave propagation by directly solving Maxwell's equations, theoretically providing the highest accuracy. 2) Ray tracing: A simplified method that simplifies electromagnetic waves into rays to simulate their reflection, transmission, diffraction, etc. in the environment to calculate signal strength distribution. 3) Neural network-based methods: These methods use deep learning models to directly learn the mapping relationship from environmental layout (such as floor plan) and access point (AP) location to radio map. For example, RadioUNet uses a U-Net network structure for end-to-end generation; RME-GAN uses a generative adversarial network (GAN) for construction; SIP2Net introduces asymmetric convolution and empty spatial pyramid pooling on the basis of U-Net to improve performance. These methods usually take environmental geometric information as input and output pixel-level signal strength maps.

[0006] However, the above prior art has the following problems in practical application: The method based on the ranging model: the indoor environment is extremely complex, and there are serious multipath propagation, signal penetration and diffraction effects, which cause the relationship between signal strength and distance to no longer meet the simple statistical model, thereby generating huge ranging error, and the positioning accuracy cannot be guaranteed.

[0007] The method based on the fingerprint library: the construction and maintenance of the fingerprint library are extremely high in cost, which needs to consume a lot of manpower and time for carpet signal collection. In addition, any slight change of the indoor environment (such as furniture moving, opening and closing doors) can cause the fingerprint library to fail, and frequent updating is needed, which has poor scalability.

[0008] The method based on simulation: 1) high calculation delay: although the full-wave electromagnetic solver is accurate, its memory and calculation demand increases exponentially, which is only suitable for small-scale simulation, and is not practical for room or floor level deployment. Although the ray tracing is accelerated, it still needs minutes of calculation time when dealing with complex multipath, which cannot meet the real-time updating demand after the environment changes. 2) dependent on sparse measurement: most of the existing neural network methods need to rely on partial real, sparse signal measurement data for training or calibration, which limits their application in the "zero measurement" scene. 3) poor physical consistency and insufficient generalization ability: the existing measurement-free neural network methods (such as RadioUNet) are mostly designed for outdoor scenes, which usually assume that the environmental materials are uniform. This assumption is completely not established in the indoor environment, because the indoor environment is full of heterogeneous materials such as concrete, glass and wood. The mismatch between this model design and the actual application conditions leads to low accuracy of the generated radio map at the boundary of walls, doors and windows, and cannot accurately capture the signal mutation caused by material differences. The purely data-driven model is difficult to learn the material-sensitive propagation law, resulting in poor generalization performance in new building layouts or material configurations. SUMMARY

[0009] In order to solve the above problems in the prior art, the present application provides an indoor radio map construction method and device based on a physically enhanced diffusion model. The technical problems to be solved by the present application are solved by the following technical solutions: In a first aspect, the embodiments of the present application provide an indoor radio map construction method based on a physically enhanced diffusion model, which comprises: constructing an initial discontinuity physical prior according to the geometric information and building material information of the indoor environment; the initial discontinuity physical prior includes a diffraction point set and a transmission boundary point set; pruning the diffraction point set in the initial discontinuity physical prior to update the discontinuity physical prior; generate a discontinuity physical prior map according to the indoor radio access points and the updated discontinuity physical prior; construct a multi-modal physical prior; wherein the multi-modal physical prior comprises coordinate positions of the indoor radio access points, a reflection coefficient map and a transmission coefficient map characterizing building materials at each position, and the discontinuity physical prior map; After processing the multi-modal physical prior through a multi-modal fusion encoder, input the processing result as a model input condition into the trained decoupled diffusion model to generate an indoor radio map.

[0010] In an embodiment of the present application, according to the geometric information and building material information of the indoor environment, an initial discontinuity physical prior is constructed, comprising: By a geometric neighborhood analysis algorithm, locate the effective corner of diffraction according to the geometric information of the indoor environment, and form a diffraction point set from all effective corners; According to the building material information of the indoor environment, form a transmission coefficient map, and locate the region of strong transmission by thresholding the transmission coefficient map, and form a transmission boundary point set from the boundary points of all strong transmission regions.

[0011] In an embodiment of the present application, the diffraction point set in the initial discontinuity physical prior is pruned, comprising: According to the geometric information of the indoor environment, remove the diffraction points in the diffraction point set that cannot form an effective incident shadow area, and remove the diffraction points in the diffraction point set whose two-sided regions are geometrically connected and do not constitute a signal mutation boundary.

[0012] In an embodiment of the present application, the process of generating the discontinuity physical prior map comprises: Take the indoor radio access point as the starting point, and take the direction from the indoor radio access point to each diffraction point and transmission boundary point in the updated discontinuity physical prior as the extension direction to form several rays; Remove the part between the indoor radio access point and each diffraction point and transmission boundary point in the updated discontinuity physical prior in each ray, and generate a discontinuity physical prior map according to the remaining part of each ray.

[0013] In an embodiment of the present application, the multi-modal fusion encoder is a convolutional layer with a 1x1 convolution kernel.

[0014] In an embodiment of the present application, the decoupled diffusion model comprises a conditional encoding module, a noise feature extraction module, and a dual target recovery and fusion module; wherein, The conditional encoding module is used to encode the model input condition obtained by processing through the multi-modal fusion encoder by an independent conditional encoder to obtain conditional features; a noise feature extraction module configured to perform multi-scale feature extraction on the noisy indoor radio map of the current iteration by a U-Net encoder to obtain noise features; a dual target recovery and fusion module including two parallel U-Net decoders configured to fuse the noise features and the conditional features by a cross-attention mechanism based on the two parallel U-Net decoders, and recover two independent target outputs from the two parallel U-Net decoders based on the fused features, wherein one U-Net decoder recovers a prediction of the current noise component, and the other U-Net decoder recovers a direct prediction of the noise-free indoor radio map, so as to calculate and generate the noisy indoor radio map required for the next iteration by the inverse diffusion sampling algorithm based on the two target outputs recovered by the dual target recovery and fusion module until a clear indoor radio map is finally recovered.

[0015] In an embodiment of the present application, the indoor radio map dataset used in the decoupled diffusion model training process includes an antenna position generalized indoor radio map dataset and a zero-shot layout generalized indoor radio map dataset.

[0016] In a second aspect, an embodiment of the present application provides an indoor radio map construction device based on a physically enhanced diffusion model, which comprises: a first construction module configured to construct an initial discontinuity physical prior according to geometric information and building material information of an indoor environment; the initial discontinuity physical prior includes a set of diffraction points and a set of transmission boundary points; a pruning module configured to prune the set of diffraction points in the initial discontinuity physical prior to update the discontinuity physical prior; a first generation module configured to generate a discontinuity physical prior map according to an indoor radio access point and the updated discontinuity physical prior; a second construction module configured to construct a multi-modal physical prior; wherein the multi-modal physical prior includes a coordinate position of an indoor radio access point, a reflection coefficient map and a transmission coefficient map representing building materials at each position, and the discontinuity physical prior map; a second generation module configured to input the multi-modal physical prior processed by a multi-modal fusion encoder as a model input condition into the trained decoupled diffusion model to generate an indoor radio map.

[0017] The present application has the following beneficial effects: The indoor radio map construction method based on the physical enhanced diffusion model overcomes the problems of high calculation delay, dependence on field measurement and insufficient modeling of complex indoor environments in existing indoor radio map construction technologies, and is a method for constructing an indoor radio map without field measurement, which can quickly generate a high-fidelity indoor radio map, and has the following advantages: by taking the material electromagnetic properties and key physical structures (diffraction points and transmission boundary points) as a priori, the indoor radio map generated by the present application can highly realistically restore the propagation details of signals in complex indoor environments, especially in boundary areas where signals change dramatically, and the accuracy is much higher than that of existing technologies; the method disclosed by the present application does not require any field signal measurement, and can generate a high-precision indoor radio map only by relying on the transmission coefficient map, the reflection coefficient map, the single access point AP position and the discontinuity physical priori map, thereby greatly reducing the deployment cost and time and supporting rapid response to dynamic changes in the environment; the decoupled diffusion model adopted by the present application can effectively generalize to new indoor layouts and access point AP positions because the model learns the universal electromagnetic propagation physical law rather than the data distribution of a specific scene, and therefore has strong robustness; experiments prove that the indoor radio map generated by the present application as a fingerprint library can achieve an indoor positioning accuracy of less than 10 meters, thereby providing a solid technical foundation for various downstream applications requiring accurate position information.

[0018] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of a method for constructing an indoor radio map based on a physical enhanced diffusion model according to an embodiment of the present application; Figure 2 is a schematic diagram of generating a discontinuity physical priori map according to an embodiment of the present application; Figure 3 is a schematic diagram of a network model suitable for indoor radio map generation according to an embodiment of the present application; Figure 4 is a process diagram of iterative denoising using a decoupled diffusion model according to an embodiment of the present application; Figure 5 is a comparison diagram of visualized generation results of each method in an antenna position generalization scenario according to an embodiment of the present application; Figure 6 is a comparison diagram of visualized generation results of each method in a zero-sample layout generalization scenario according to an embodiment of the present application; Figure 7 is a structural diagram of an indoor radio map construction device based on a physical enhanced diffusion model according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] The application will be described in further detail below with reference to specific embodiments, but the embodiments of the application are not limited thereto.

[0021] In order to overcome the defects of high calculation delay, dependence on field measurement and insufficient modeling of complex indoor environment in existing radio map construction technology, the present application designs a radio map construction method for indoor environment which does not need field measurement, can be quickly generated and has high fidelity. In order to achieve this goal, the inventor found that the following problems need to be considered in the design process: first, the method needs to adapt to the material heterogeneity of the indoor environment, by explicitly incorporating the electromagnetic properties (such as reflection and transmission coefficients) of different materials such as walls, doors and windows into the model, breaking the "uniform medium" assumption commonly existing in traditional methods, thereby improving the physical authenticity; second, the method must accurately model the discontinuity of the signal, accurately capture the signal strength discontinuity caused by wall corner diffraction, door and window transmission and other physical phenomena, and solve the problem that the existing convolution network is difficult to handle such non-stationary field distribution due to its smooth prior; third, the method should ensure physical consistency and high generalization ability, ensuring that the generated radio map not only conforms to the basic laws of electromagnetic wave propagation, but also can effectively generalize to indoor layouts and material configurations that have never been seen; finally, the method needs to support real-time updating, when the environment changes, it can quickly regenerate the map by simply updating the physical parameters, meeting the needs of dynamic scenes. Finally, through the consideration of the above problems, the present application proposes an indoor radio map construction method and device based on a physically enhanced diffusion model.

[0022] In a first aspect, see Figure 1 The embodiment of the present application provides a radio map construction method for indoor environment based on a physically enhanced diffusion model, which specifically comprises the following steps: S10, constructing an initial discontinuity physical prior according to the geometric information and building material information of the indoor environment; the initial discontinuity physical prior includes a diffraction point set and a transmission boundary point set.

[0023] Unlike existing indoor radio map construction methods that only take geometric layout as input, the present application recognizes that non-stationary discontinuities of indoor signals are the main source of prediction errors. These non-stationary discontinuities are dominated by specific electromagnetic propagation phenomena such as diffraction and strong transmission, which are difficult to capture by standard convolutional networks due to their inherent smoothness prior. Therefore, the present application proposes to explicitly encode these discontinuities prior as the guiding information of the model. Specifically, embodiments of the present application construct an initial discontinuity physical prior according to the geometric information and building material information of the indoor environment, including: locating the effective corner of diffraction according to the geometric information of the indoor environment by a geometric neighborhood analysis algorithm, and forming a diffraction point set from all effective corners; forming a transmission coefficient map according to the building material information of the indoor environment, thresholding the transmission coefficient map, locating the region where strong transmission occurs, and forming a transmission boundary point set from the boundary points of all strong transmission regions.

[0024] S20, pruning the diffraction point set in the initial discontinuity physical prior to update the discontinuity physical prior.

[0025] After identifying the discontinuity physical prior, the present application prunes the discontinuity physical prior to filter out redundant and invalid prior information, ensuring that the guiding information of the input model has high relevance and low noise. Specifically, embodiments of the present application prune the diffraction point set in the initial discontinuity physical prior, including: according to the geometric information of the indoor environment, eliminating diffraction points in the diffraction point set that cannot form an effective incident shadow area, and eliminating diffraction points in the diffraction point set whose two-sided regions are geometrically connected and do not constitute a signal discontinuity boundary. More specifically: The pruning process based on the effective incident shadow area in the present application follows the basic principles of geometric diffraction theory, and the diffraction effect of a diffraction point can only be ignored when it can form an effective incident shadow area. For this purpose, the present application first calculates the transition profile of the line-of-sight region and the non-line-of-sight region from the indoor radio access point AP, i.e. the line-of-sight boundary, to determine whether each diffraction point can form an effective obstruction with respect to the indoor radio access point AP. If the diffraction point and its obstacle surface cannot form an effective incident shadow area, i.e. it is located in the full view area of the indoor radio access point AP, then the diffraction point is considered to be an invalid diffraction point and is removed.

[0026] To further refine the discontinuity physical prior, the present application analyzes the path loss consistency on both sides of the diffraction point through the geometric information of the indoor environment to remove redundant diffraction points. If the regions on both sides of a corner of the indoor environment experience similar penetration loss with respect to the indoor radio access point AP, i.e. the path loss change of the signal propagating to both sides of the corner is smooth, then the diffraction point will not constitute a signal discontinuity boundary, and such diffraction points are determined to be redundant and removed from the diffraction point set.

[0027] S30, generating a discontinuity physical prior map according to the indoor radio access point and the updated discontinuity physical prior.

[0028] The process of generating the discontinuity physical prior map in the embodiment of the application comprises: taking the indoor radio access point as a starting point, and taking the directions of each diffraction point and transmission boundary point in the updated discontinuity physical prior from the indoor radio access point as an extension direction to form a plurality of rays; removing the part between the indoor radio access point and each diffraction point and transmission boundary point in each ray, and generating the discontinuity physical prior map according to the remaining part of each ray. Figure 2 As shown in the figure, the blue points are transmission boundary points, the green points are diffraction points, the rays include the rays passing through the indoor radio access point and each diffraction point as shown by the green lines, and the rays passing through the indoor radio access point and each transmission boundary point as shown by the red lines, and the removed part of each ray is shown by the dashed lines, that is, the line segment between the indoor radio access point and each diffraction point and transmission boundary point in each ray is removed.

[0029] S40, constructing a multi-modal physical prior; wherein the multi-modal physical prior comprises the coordinate position of the indoor radio access point, the reflection coefficient map and the transmission coefficient map representing the building material at each position, and the discontinuity physical prior map.

[0030] In the physical condition input construction of the model, the embodiment of the application takes the physical information under four modes as prior information, which considers the coordinate position of the indoor radio access point, the reflection coefficient map and the transmission coefficient map representing the building material at each position, and more importantly, considers the discontinuity physical prior map, takes the most critical and most error-prone signal mutation part as a prior condition, so that the model can perceive where the signal mutation may occur in the generation process, thereby avoiding the over-smoothing problem caused by the existing method, and significantly improving the modeling accuracy of physical details.

[0031] S50, processing the multi-modal physical prior through a multi-modal fusion encoder, and inputting the processing result into the trained decoupled diffusion model as a model input condition to generate an indoor radio map.

[0032] The embodiment of the application proposes a network model suitable for indoor radio map generation, as shown in the figure. Figure 3 The network model comprises a multi-modal fusion encoder and a decoupled diffusion model. The multi-modal fusion encoder is introduced because the decoupled diffusion model expects to receive a three-channel image as a conditional input, and the multi-modal physical prior constitutes a four-channel image. The application converts the multi-modal physical prior into a data form matching the input of the decoupled diffusion model by designing a multi-modal fusion encoder compatible with the decoupled diffusion model.

[0033] The multi-modal fusion encoder in the embodiment of the application is a convolution layer with a convolution kernel of 1*1, Figure 3 The leftmost Conv in the embodiment of the application is a lightweight 1*1 convolution operation, which realizes a learnable fully connected linear transformation of all input channels at each pixel position, and learns to optimally combine, weight and fuse the four kinds of heterogeneous information of reflection, transmission, AP proximity and discontinuity prior, and finally encodes them into three channel features with more intensive and rich information.

[0034] The decoupled diffusion model in the embodiment of the application includes a conditional encoding module, a noise feature extraction module and a dual target recovery and fusion module. Figure 3 The conditional encoding module is configured to encode the model input condition (Cond) processed by the multi-modal fusion encoder to obtain a condition feature. Figure 3 The noise feature extraction module is configured to perform multi-scale feature extraction on the current iteration of the noisy indoor radio map (Input) by a U-Net encoder (UNet Encoder) to obtain a noise feature. The dual target recovery and fusion module includes two parallel U-Net decoders (UNet Decoder1 and UNet Decoder2), which are configured to fuse the noise feature and the condition feature based on the two parallel U-Net decoders through a cross-attention mechanism, and recover two independent target outputs from the two parallel U-Net decoders according to the fused feature, wherein one U-Net decoder recovers a prediction of the current noise component, and the other U-Net decoder recovers a direct prediction of the noise-free indoor radio map, so as to calculate and generate the noisy indoor radio map required for the next iteration by the reverse diffusion sampling algorithm according to the two target outputs recovered by the dual target recovery and fusion module, until a clear indoor radio map is finally recovered.

[0035] The model input condition of the decoupled diffusion model is defined as a set represents the coordinate position of the indoor radio access point, represents the reflection coefficient map of the building material at each position, represents the transmission coefficient map of the building material at each position, represents a discontinuity physical prior map, which is generated by the discontinuity physical prior. The decoupled diffusion model in the embodiment of the application is capable of learning a mapping relationship from the input condition to the estimated indoor radio map , i.e. The decoupled diffusion model In the training data set , N represents the total number of samples in the training data set, and the indoor radio map data set used as the training data set in the training process includes: an indoor radio map data set generalized by antenna position and an indoor radio map data set generalized by zero-sample layout. Therefore, the optimization target of the present application can be defined as finding a set of optimal network parameters to minimize the mean square error between the network prediction of the indoor radio map and the real indoor radio map. The optimization problem is expressed as follows: ; wherein, represents the model input condition corresponding to the th training sample in the training process, represents the indoor radio map corresponding to the th training sample in the training process. The objective function minimizes the prediction error, so that the decoupled diffusion model learns the complex electromagnetic wave propagation law in the room, thereby generating an indoor radio map highly consistent with the physical law under the condition of only given AP position and building material attributes.

[0036] Here, it is assumed that a high-fidelity indoor radio map is constructed in a two-dimensional indoor area , which is discretized into a uniform grid , and the indoor radio map is represented as a received signal strength (RSSI) matrix .

[0037] The decoupled diffusion model of the embodiment of the present application is shown in Figure 4 , which recovers a high-fidelity radio map with multi-modal physical information as the model condition input from a noisy image through an iterative denoising process, Figure 4 from left to right, is a random Gaussian noise image, , is a noisy indoor radio map in the iterative denoising process, is the recovered generated high-fidelity radio map. At each step of denoising, the model can dynamically focus on the most important physical information at the current spatial position, for example, when processing the pixels near the corner of the wall, the attention is focused on the input diffraction prior image, thereby generating an indoor radio map highly consistent with the physical law. This design explicitly embeds the physical constraints of electromagnetic propagation into the learning process of the generation model, improving the modeling accuracy of physical details.

[0038] In order to verify the effectiveness of the indoor radio map construction method based on the physically enhanced diffusion model provided by the embodiment of the present application, the following experiments are carried out to verify the effectiveness.

[0039] The proposed method is comprehensively evaluated using a publicly available indoor radio map dataset. This indoor radio map dataset is generated by a ray-tracing method, contains 25 different indoor layouts, and provides a physically realistic indoor radio map that considers multiple reflections, transmissions, and diffractions, at a 3.5 GHz frequency band, ignoring the differences in antenna radiation patterns to isolate the environmental-induced propagation effects. Each environment is discretized into a 0.25-meter grid, and the simulation considers up to 8 reflections, 10 transmissions, and 2 diffractions.

[0040] To systematically evaluate the generalization capability of the model, the evaluation protocol is divided into two scenarios: Antenna location generalization (ALG): The training dataset contains all 25 scene layouts, but only 45 out of 50 AP locations are used for each layout; the test dataset uses the remaining 5 never-seen-before AP locations in the same layout to evaluate the model's adaptability to new deployments.

[0041] Zero-shot layout generalization (ZLG): The training dataset contains 20 scene layouts, each with 50 AP locations, while the test dataset contains 5 completely new scenes that the model has never seen during training, to evaluate the model's robustness in unknown environments.

[0042] In quantitative evaluation, multi-dimensional indicators are used, including the root mean square error (RMSE) and the peak signal-to-noise ratio (PSNR) to measure pixel-level error. In addition, to better capture the critical perceptual quality and detail distribution in indoor radio maps, the learned perceptual image patch similarity (LPIPS) and Fréchet inception distance (FID) are introduced as additional evaluation indicators. LPIPS is used to measure the perceptual similarity of generated maps and real maps in the deep feature space, while FID evaluates the consistency of the overall distribution, which is crucial for modeling the fine-grained changes in signals caused by multipath effects.

[0043] All experiments are run on an NVIDIA GeForce RTX 4090 running the Ubuntu operating system. To comprehensively evaluate the effectiveness of the proposed method, denoted as iRadioDiff, the present invention compares it with three representative deep learning architectures used for radio map construction, including: RadioUNet (CNN architecture), RME-GAN (generative adversarial network), and SIP2Net (enhanced U-Net architecture based on asymmetric convolution and ASPP).

[0044]

[0045] As can be seen from Table 1: in the antenna position generalization ALG scene, the RMSE of the present application is as low as 6.357, which is reduced by 32% compared with the suboptimal RadioUNet (9.349), and the precision realizes a qualitative leap. At the same time, the PSNR is as high as 32.24, and the LPIPS and FID are also far ahead with the scores of 0.2742 and 145.2 respectively, which shows that the map generated by the present application not only has lower pixel-level error, but also has closer overall structure and detail distribution to the physical reality. In the more challenging zero-shot layout generalization (ZLG) scene, the present application still maintains strong performance. Although facing a new environment, the RMSE is still maintained at 7.010, and the PSNR is 31.45, which are better than all the comparison methods.

[0046] The experimental results show that the iRadioDiff method proposed in the present application almost realizes the best performance in all evaluation indexes of the two generalization scenes, and completely surpasses the existing deep learning methods.

[0047] Further, the effectiveness of the iRadioDiff method proposed in the present application in the positioning task is verified. The KNN algorithm with K=5 is used for positioning test in the positioning verification, and the positioning task results are as shown in Table 2.

[0048] Table 2 Positioning task results

[0049] As can be seen from Table 2: in the positioning task, the iRadioDiff proposed in the present application is the only method that can stably realize sub-10-meter-level positioning accuracy, and the positioning errors in the ALG and ZLG scenes are 7.860 meters and 8.530 meters respectively, which shows that the present application can accurately infer the signal distribution of new AP positions in known scenes, the physical law learned by the present application has strong migratory property, can be effectively applied to a new building environment, and has outstanding generalization ability, and the radio map generated by the present application has outstanding value in practical application.

[0050] In addition, Figure 5 and Figure 6 respectively illustrate the visual generation results of each method in the antenna position generalization scene and the zero-shot layout generalization scene, Figure 5 and Figure 6The visualization generation results further emphasize the advantages of the iRadioDiff method proposed in the present application: it can more effectively capture the fine structure details and signal discontinuities caused by multipath propagation, diffraction and transmission, thereby achieving a more clear and physically detailed reconstruction of the radio map. Especially in areas such as corners and doorways where signal strength changes dramatically, the present application can accurately model these fine-grained changes, while traditional methods such as RadioUNet exhibit obvious oversmoothing phenomena and fail to successfully capture these key physical information.

[0051] To verify the effectiveness of the "physical information prior" proposed in the present application, an ablation study was conducted. In the ablation study, the method of the present application using the complete physical information prior is marked as "w / Physics", all physical-related prior inputs (i.e. discontinuity physical prior map) are removed, and the model only receives the reflectance coefficient map, the transmission coefficient map and the ablation experiment version of the AP position is marked as "w / o Physics", and the ablation study results are shown in Table 3.

[0052] Table 3 Ablation study of physical information prior

[0053] As can be seen from Table 3, when the physical prior is removed ("w / o Physics"), the performance indicators of the model all show a significant decline, including the accuracy of indoor radio map construction and positioning accuracy. For example, in the ALG task, the RMSE worsens from 6.357 to 9.619, and the positioning error also increases accordingly. This result strongly proves that the strategy of injecting material electromagnetic properties, diffraction, transmission boundary and other physical priors into the diffusion model proposed in the present application is the key to achieving high fidelity and strong generalization ability.

[0054] The above experimental results fully verify that the physically enhanced diffusion model of the present application not only greatly improves the reconstruction accuracy of the indoor radio map, but also maintains its excellence in complex environments such as unknown antenna positions and completely new indoor layouts, making it a powerful tool for high-precision positioning and network optimization in the next generation of communication networks.

[0055] In summary, the indoor radio map construction method based on the physical enhanced diffusion model proposed in the embodiments of the present application overcomes the problems of high calculation delay, dependence on field measurement and insufficient modeling of complex indoor environments in existing indoor radio map construction technologies, and is a fast generation, high-fidelity indoor radio map construction method without field measurement, which has the following significant advantages: By taking the material electromagnetic properties and key physical structures (diffraction points, transmission boundary points) as a priori, the indoor radio map generated by the present application can highly realistically restore the propagation details of signals in complex indoor environments, especially in the boundary area where the signal changes dramatically, and the accuracy is much higher than that of the prior art; The method proposed in the present application does not require any field signal measurement at all, and only relies on the transmission coefficient map, the reflection coefficient map, the single access point AP position and the discontinuity physical priori map to generate a high-precision indoor radio map, which greatly reduces the deployment cost and time, and supports fast response to dynamic changes in the environment; The decoupled diffusion model used in the present application learns the universal electromagnetic propagation physical law, rather than the data distribution of a specific scene, so it can effectively generalize to new indoor layouts and access point AP positions, and has strong robustness; Experiments show that using the indoor radio map generated by the present application as a fingerprint library can achieve sub-10-meter indoor positioning accuracy, providing a solid technical foundation for various downstream applications that require accurate position information.

[0056] In a second aspect, referring to Figure 7 The embodiments of the present application provide an indoor radio map construction device based on a physical enhanced diffusion model, which comprises: A first construction module is configured to construct an initial discontinuity physical priori according to the geometric information and the building material information of the indoor environment; the initial discontinuity physical priori comprises a diffraction point set and a transmission boundary point set. A pruning module is configured to prune the diffraction point set in the initial discontinuity physical priori to update the discontinuity physical priori. A first generation module is configured to generate a discontinuity physical priori map according to the indoor radio access point and the updated discontinuity physical priori. A second construction module is configured to construct a multi-modal physical priori; wherein the multi-modal physical priori comprises the coordinate position of the indoor radio access point, the reflection coefficient map and the transmission coefficient map representing the building material at each position, and the discontinuity physical priori map. A second generation module is configured to input the multi-modal physical priori into a multi-modal fusion encoder for processing, and then input the processing result as a model input condition into a trained decoupled diffusion model to generate an indoor radio map.

[0057] For the device embodiments of the second aspect, since they are basically similar to the method embodiments of the first aspect, the description is relatively simple, and the relevant parts refer to the part of the description of the method embodiments of the first aspect.

[0058] In the description of the present application, it should be understood that the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specifically limited.

[0059] Although the present application is described herein in conjunction with various embodiments, other variations of the disclosed embodiments can be understood and implemented by those skilled in the art with reference to the specification and drawings. In the specification, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude a plurality. Some measures are described in mutually different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0060] The above is a further detailed description of the present application in conjunction with specific preferred embodiments, and cannot be considered as limiting the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the protection scope of the present application.

Claims

1. A method for constructing indoor radio maps based on a physical augmentation diffusion model, characterized in that, The indoor radio map construction method includes: Based on the geometric information of the indoor environment and the information of building materials, an initial discontinuous physical prior is constructed; the initial discontinuous physical prior includes the diffraction point set and the transmission boundary point set. The diffraction point set in the initial discontinuous physical prior is pruned to update the discontinuous physical prior; Generate a discontinuous physical prior map based on the indoor radio access point and the updated discontinuous physical prior. Construct a multimodal physical prior; the multimodal physical prior includes the coordinate location of the indoor radio access point, the reflection coefficient map and transmission coefficient map characterizing the building materials at each location, and the discontinuous physical prior map; The multimodal physical priors are processed by a multimodal fusion encoder, and the processing result is used as the model input condition to input into the trained decoupled diffusion model to generate an indoor radio map.

2. The indoor radio map construction method based on the physical augmentation diffusion model according to claim 1, characterized in that, Based on the geometric information of the indoor environment and the information of building materials, an initial discontinuity physical prior is constructed, including: Using a geometric neighborhood analysis algorithm, the effective corners where diffraction occurs are located based on the geometric information of the indoor environment, and the set of diffraction points is formed by all effective corners. A transmission coefficient map is generated based on the building material information of the indoor environment. The transmission coefficient map is then thresholded to locate areas with strong transmission. A transmission boundary point set is formed by the boundary points of all areas with strong transmission.

3. The method for constructing indoor radio maps based on a physically enhanced diffusion model according to claim 1, characterized in that, Pruning the diffraction point set in the initial discontinuous physical prior includes: Based on the geometric information of the indoor environment, diffraction points that cannot form an effective incident shadow area in the diffraction point cluster are eliminated, as are diffraction points whose regions on both sides of the diffraction point cluster are geometrically connected and do not constitute a signal abrupt boundary.

4. The method for constructing indoor radio maps based on a physical augmentation diffusion model according to claim 1, characterized in that, The process of generating a discontinuous physical prior map includes: Taking the indoor radio access point as the starting point, and the direction from the indoor radio access point to each diffraction point and transmission boundary point in the updated discontinuous physical prior as the extension direction, several rays are formed. Remove the portion of each ray between the indoor radio access point and each diffraction point and transmission boundary point in the updated discontinuity physics prior, and generate a discontinuity physics prior map based on the remaining portion of each ray.

5. The method for constructing indoor radio maps based on a physically enhanced diffusion model according to claim 1, characterized in that, The multimodal fusion encoder is a convolutional layer with a 1×1 kernel.

6. The method for constructing indoor radio maps based on a physically enhanced diffusion model according to claim 1, characterized in that, The decoupled diffusion model includes a conditional coding module, a noise feature extraction module, and a dual-target recovery and fusion module; among which, The conditional encoding module is used to encode the model input conditions obtained by the multimodal fusion encoder through an independent conditional encoder to obtain conditional features; The noise feature extraction module is used to extract noise features from the noisy indoor radio map of the current iteration using a U-Net encoder at multiple scales. The dual-target recovery and fusion module includes two parallel U-Net decoders. Based on the two parallel U-Net decoders, noise features and conditional features are fused through a cross-attention mechanism. According to the fused features, the two parallel U-Net decoders recover two independent target outputs. One U-Net decoder recovers the prediction of the current noise component, and the other U-Net decoder recovers the direct prediction of the noise-free indoor radio map. Based on the two target outputs recovered by the dual-target recovery and fusion module, the noisy indoor radio map required for the next iteration is calculated and generated through a back-diffusion sampling algorithm until a clear indoor radio map is finally recovered.

7. The method for constructing indoor radio maps based on a physical augmentation diffusion model according to claim 1, characterized in that, The indoor radio map datasets used in the training of the decoupled diffusion model include an indoor radio map dataset generalized to antenna location and an indoor radio map dataset generalized to zero-sample layout.

8. An indoor radio map construction device based on a physical augmentation diffusion model, characterized in that, The indoor radio map building device includes: The first construction module is used to construct the initial discontinuous physical prior based on the geometric information of the indoor environment and the building material information; the initial discontinuous physical prior includes the diffraction point set and the transmission boundary point set; The pruning module is used to prune the diffraction point set in the initial discontinuous physical prior in order to update the discontinuous physical prior. The first generation module is used to generate a discontinuous physical prior map based on the indoor radio access point and the updated discontinuous physical prior. The second building module is used to construct multimodal physical priors; among which, multimodal physical priors include the coordinate location of the indoor radio access point, the reflection coefficient map and transmission coefficient map characterizing the building materials at each location, and the discontinuous physical prior map; The second generation module is used to process the multimodal physical priors through a multimodal fusion encoder, and then use the processing results as input conditions to the trained decoupled diffusion model to generate an indoor radio map.

Images