Search

CN-121984623-A - Low-altitude XL-MIMO channel generation and base model application method based on digital twin

CN121984623ACN 121984623 ACN121984623 ACN 121984623ACN-121984623-A

Abstract

The invention discloses a digital twin-based low-altitude XL-MIMO channel generation and base model application method, which is used for constructing a 3D static environment endowed with electromagnetic properties, simulating urban road following and suburban straight line combining height-changing flight tracks, generating a 3D dynamic scene and constructing a channel data set containing multipath components. XL-MIMO generated base model based on denoising diffusion hidden model is designed and pre-trained, effective noise is predicted by random mask and forward noise injection network, and complete channel is reconstructed from noise by means of conditional DDIM reverse sampling. The method comprises the steps of freezing a plurality of convertors layers at the front end of a model, updating only a small number of layers at the rear end, fusing a multi-time characteristic module and a lightweight task head, and aiming at different tasks of channel estimation, user classification and 3D positioning, respectively adopting a corresponding processing mode and the task head to realize near-far field state discrimination and high-precision three-dimensional coordinate prediction of an aircraft, thereby greatly improving the robustness and accuracy of low-altitude communication under sparse observation.

Inventors

  • HAN YU
  • LI MENGYUAN
  • JIN SHI

Assignees

  • 东南大学

Dates

Publication Date
20260505
Application Date
20260205

Claims (7)

  1. 1. The low-altitude XL-MIMO channel generation and base model application method based on digital twinning is characterized by comprising the following steps of: step 1, acquiring 3D geography and material attribute information of a target area, constructing a 3D static environment, simulating a 3D flight track of a low-altitude aircraft in the target area to form a 3D dynamic scene, and constructing an XL-MIMO channel data set based on digital twinning based on the 3D static environment and the dynamic scene by utilizing a ray tracing technology; Step 2, constructing a generated base model based on a denoising diffusion hidden model by taking a transform architecture as a backbone network, pre-training the generated base model by utilizing the XL-MIMO channel data set constructed in the step 1 to enable the generated base model to learn to recover a complete channel from noise and partial observation channels, wherein the backbone network comprises a plurality of transform layers, each transform layer comprises a plurality of mask multi-head attention modules, the input of a first transform layer is embedded in a blocking manner through a convolution layer, and the mask multi-head attention modules of the first transform layer are input after position codes and time embedments are overlapped, and the output of the previous transform layer is overlapped with the position codes and the time embedments and then is input to the next transform layer; step 3, freezing parameters of the front B transducer layers of the backbone network on the basis of the pre-trained generated base model, designing corresponding task heads according to downstream wireless communication tasks, and training and updating the parameters of the task heads and the remaining L transducer layers; And 4, designing a denoising head based on a U-Net architecture for denoising the received signal aiming at a noise-containing pilot frequency scene, then carrying out channel reconstruction by using a backbone network, classifying the downstream wireless communication task into low-altitude aircrafts, designing a multi-time feature fusion module, extracting features of different time steps in the diffusion process, judging whether the low-altitude aircrafts are in a near field or a far field through the backbone network and the classification head, and mapping the fused high-dimensional features into 3D space coordinates of the low-altitude aircrafts by using the backbone network and the regression head when the downstream wireless communication task is 3D positioning of the low-altitude aircrafts.
  2. 2. The digital twinning-based low-altitude XL-MIMO channel generation and base model application method of claim 1, wherein the specific process of step 1 is as follows: extracting geographic information of static objects in a target area by utilizing OpenStreetMap, wherein the geographic information comprises building outlines, heights and road networks, and endowing corresponding electromagnetic material properties for different objects according to ITU standards to construct a 3D static environment; simulating a horizontal movement track by utilizing a SUMO frame based on the extracted road network aiming at the urban scene, and generating a change track of a height dimension by combining a sine wave or a limited random walk model; the method comprises the steps of randomly sampling speed on a horizontal plane and simulating an approximate straight-line flight track aiming at suburb scenes, generating a change track of a height dimension by combining a sine wave or a limited random walk model, mapping generated track coordinates from a local coordinate system back to a WGS84 coordinate system, and projecting the generated track coordinates to an engineering coordinate system to be aligned with a 3D static environment; channel generation is carried out by utilizing Sionna RT ray tracing technology, namely, when the low-altitude aircraft is in a near-field region, an antenna-by-antenna generation mode is adopted, and when the low-altitude aircraft is in a far-field region, a synthetic array mode is adopted, and the path length, the complex gain, the departure angle and the arrival angle of multipath components are calculated to generate channel impulse response, wherein the channel impulse response is specifically as follows: Base station configuration Uniform plane array of root antenna, setting central coordinate of array as First, the The center coordinates of the root antenna are The low-altitude aircraft coordinates are Calculating Rayleigh distance The method comprises the following steps: , Wherein, the , The number of antennas in the vertical and horizontal directions respectively, The antenna spacing is in the vertical and horizontal directions respectively, As a function of the wavelength(s), , ; When (when) When the low-altitude aircraft is in the near-field region, for each pair of transceiving links Calculation of Strip multipath component, 1 Receiving electric field of strip path The method comprises the following steps: , Wherein, the As an initial component of the electric field, For the geometric length of the path, For the gain of the transmitting end, In order for the receiving-end gain to be high, Respectively the first The azimuth and pitch angles of the strip path, Is the first Fresnel coefficients of sub-reflection or diffraction, Representing the number of path interactions; the channel response is: , Wherein, the Is the first The channel response of the root antenna, For the path gain to be a function of the path gain, , In units of imaginary numbers, Is the path length; When (when) When the low-altitude aircraft is in a far-field area, the low-altitude aircraft is calculated To the point of Including departure angle and arrival angle, and using steering vectors A full array channel is synthesized.
  3. 3. The digital twinning-based low-altitude XL-MIMO channel generation and base model application method of claim 1, wherein in the step 2, the pre-training process of the generated base model is as follows: forward diffusion process, namely, the complete channel generated in the step 1 is processed Masking operation to obtain partial observation channel And at Gradually adding Gaussian noise to the first part Obtaining noise channels ; Noise prediction is to 、 Mask code Block coding and position coding, time coding input backbone network, predictive first Effective noise of steps The pre-trained objective function is to minimize the mean square error between the predicted noise and the true additive noise The method comprises the following steps: , Wherein, the The result of the calculation of the expectation is indicated, As the true value of the noise is, Is prediction noise; Deterministic back sampling to partially observe the channel Under the condition, the complete channel is recovered from the random noise by gradual denoising through a deterministic reverse updating formula by using the predicted noise parameterized posterior distribution ; Given prediction noise From the first using deterministic reverse update formula Step-by-step derivation to the first Status of step : , Wherein, the Is that The channel of the moment of time, Is that The channel of the moment of time, For a complete channel, , , , Is that Noise adding parameters at the moment.
  4. 4. The digital twinning-based low-altitude XL-MIMO channel generation and base model application method according to claim 1, wherein in step 3, the parameter set defining the pre-trained generated base model is Setting parameters of B converters layers before backbone network The learning rate of (2) is 0, the parameters are kept unchanged, and the remaining L transducer layer parameters The learning rate of (2) is Setting the learning rate of the downstream task head parameters as , 。
  5. 5. The method for generating and applying a base model based on digital twin low-altitude XL-MIMO channel according to claim 4, wherein in step 4, when the downstream wireless communication task is channel estimation, a denoising head based on U-Net architecture is designed as a front of backbone network, the received noisy pilot signal and mask matrix are channel spliced as an input of the denoising head, and the denoised partial channel is input by utilizing the advanced feature of the denoising head Will be As the input of backbone network, generating complete channel by extrapolation of deterministic reverse sampling process; when the downstream wireless communication task is 3D positioning of the low-altitude aircraft, introducing a multi-time feature fusion module, and selecting Extracting feature vectors output by a backbone network in each time step, splicing the feature vectors in all time steps, adding extra time embedding and position embedding, fusing spliced features through a Transformer layer, and finally obtaining global fused features through average pooling Will be The positioning head is a regression network, the output of the positioning head comprises sine and cosine values of azimuth angle, sine and cosine values of pitch angle and distance parameters, the positioning loss function adopts a weighted summation form and comprises Euclidean distance loss aiming at angle prediction and Huber loss aiming at distance prediction and relative error loss, and the positioning loss function The following are provided: , Wherein, the The real position and the predicted position of the low-altitude aircraft are respectively, As a parameter of the weight-bearing element, For Huber loss for distance prediction, For the real distance of the low-altitude aircraft to the base station, For the predicted distance of the low-altitude aircraft to the base station, Is the maximum value of the distance from the low-altitude aircraft to the base station; When downstream wireless communication task is low-altitude aircraft classification, a multi-time feature fusion module is introduced, so that Input to a classification head, which outputs probabilities Classification loss function The following are provided: , Wherein, the Is a feature after the fusion of multiple times, Is a multi-layer perceptron module, and is characterized by that, , 。
  6. 6. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor, when executing the computer program, performs the steps of the digital twinning-based low-altitude XL-MIMO channel generation and base model application method of any one of claims 1 to 5.
  7. 7. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the digital twinning based low-altitude XL-MIMO channel generation and base model application method of any one of claims 1 to 5.

Description

Low-altitude XL-MIMO channel generation and base model application method based on digital twin Technical Field The invention relates to a low-altitude XL-MIMO channel generation and base model application method based on digital twinning, and belongs to the technical field of wireless communication. Background With the deep integration of the sixth generation mobile communication technology (6G) and Low-altitude economy (LAE) it has become a strategic place of global attention to build space-to-ground integrated stereoscopic communication networks. As a key enabling technology for realizing low-altitude wide area coverage and high-precision perception, an ultra-large-scale multi-input multi-output (XL-MIMO) system can remarkably improve the spatial resolution and the spectral efficiency of the system by means of the huge antenna array aperture, and provide reliable communication links and centimeter-level positioning services for low-altitude aircrafts such as Unmanned Aerial Vehicles (UAVs), electric vertical takeoff and landing aircrafts (eVTOL) and the like. However, unlike traditional ground communication scenarios, low-altitude communication environments have high three-dimensional dynamics, and due to the large size of XL-MIMO arrays, aircraft are often in the radiation near-field region, electromagnetic wave propagation exhibits significant spherical wavefront characteristics and spatial instability, which makes traditional far-field plane wave channel models no longer applicable, greatly increasing the complexity of channel modeling and analysis. In practical low-altitude XL-MIMO system deployment and algorithm research, high-quality and high-fidelity channel data is a basic stone for designing and evaluating various wireless communication algorithms (such as channel estimation, beam forming and user positioning). However, the current academia and industry still face a serious problem of 'data exhaustion'. On one hand, the existing open source data set (such as DeepMIMO and the like) mainly faces to ground user scenes, lacks special data aiming at a low-altitude three-dimensional airspace, is difficult to simulate the special sparse scattering environment and near-field spherical wave propagation characteristics of low altitude, and on the other hand, the acquisition of large-scale low-altitude channel data through actual measurement is high in cost, high in implementation risk and difficult to cover complex and changeable urban and suburban environments. The data loss severely restricts the development process of the low-altitude communication technology based on Artificial Intelligence (AI), and the development of a high-fidelity channel generation tool chain based on digital twinning (DIGITAL TWIN) is urgently needed to realize accurate mapping and data synthesis on a physical electromagnetic environment. Meanwhile, as artificial intelligence advances toward the Foundation Model (Foundation Model) age, learning general wireless channel characterization using a large-scale pre-training Model has become a new trend to solve complex communication tasks. Traditional AI models designed for a single task (such as only performing channel estimation or only performing positioning) tend to have weak generalization capability, and are difficult to adapt to changeable channel environments in low-altitude scenes. Although recent research has attempted to use large language model (Large language model, LLM) fine-tuning in the field of natural language processing for wireless communications, such models are large in parameter size, high in inference delay, and performance loss due to cross-modal migration is not negligible. Therefore, designing a special base model for the physical characteristics of wireless communication, pre-training on a large-scale data set to learn general channel structure characteristics, and adapting to various downstream tasks through lightweight fine tuning is a key way for improving the system efficiency and the intelligence level. However, existing wireless channel generation and pre-training methods still have a number of limitations. Conventional methods based on generation of a countermeasure network (GAN) often face the problems of Mode Collapse (Mode Collapse) and unstable training, resulting in insufficient diversity of generated channel distribution, BERT-based mask self-encoders (such as LWM) tend to generate smooth averages when processing channel data with high mask rate (i.e. very sparse pilot), which are difficult to recover high frequency details and near field phase characteristics of the channel, whereas the standard Denoising Diffusion Probability Model (DDPM) has high generation quality, but the inverse sampling process requires thousands of iterations, so that the reasoning speed is very slow, and cannot meet the severe requirements of wireless communication on low delay. In addition, most of the existing models only pay attention to data distribution of the cha