Air base station radio map generation method based on generative artificial intelligence

Abstract

The invention relates to the technical field of wireless communication, and particularly discloses an air base station radio map generation method based on generative artificial intelligence, which comprises the following steps: S1, constructing an air base station wireless channel model; s2, defining a 9D air base station radio map modeling space; s3, adopting a conditional denoising diffusion probability model to carry out offline generation type modeling on the 9D air base station radio map modeling space; and S4, based on a modeling result, constructing an air base station radio map without large-scale online measurement, and performing channel estimation of an air base station communication system. According to the method for generating the radio map of the air base station based on the generative artificial intelligence, the influence rule of the deployment state and the motion characteristic of the air base station on wireless channel distribution can be learned under the condition of a small number of collected samples, and the high-dimensional radio map is generated, so that reliable channel prior distribution is obtained.

Description

A method for generating radio maps of airborne base stations based on generative artificial intelligence Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method for generating radio maps of airborne base stations based on generative artificial intelligence. Background Technology

[0002] With the rapid development of the low-altitude economy, applications such as drones, aerial robots, low-altitude logistics transportation, low-altitude inspection, and emergency communications are constantly emerging, placing higher demands on communication coverage and data transmission capabilities in low-altitude airspace. Traditional terrestrial cellular networks struggle to provide continuous, high-quality communication services in the complex and ever-changing three-dimensional low-altitude space, and insufficient communication capabilities in low-altitude areas have become one of the key bottlenecks restricting the development of the low-altitude industry.

[0003] Aerial base stations have attracted widespread attention as a flexible and deployable communication infrastructure. Typically composed of communication payloads carried by drones, airships, or other mobile platforms, they can autonomously adjust their position and attitude in three-dimensional space to achieve functions such as enhanced local coverage, increased network capacity, and emergency communication support. Compared to traditional fixed ground base stations, aerial base stations offer greater freedom in altitude, location, and trajectory, enabling optimized network topology configurations through active maneuverability, thus playing a crucial role in the construction of low-altitude communication networks.

[0004] However, the mobility of airborne base stations also presents new technical challenges. Due to the continuous changes in platform location, attitude, and motion, the wireless channel between the airborne base station and ground users exhibits significant three-dimensional spatial variations, rapid temporal evolution, and strong non-stationarity. Channel characteristics are affected by multiple factors such as multipath propagation, obstruction, weather conditions, and ground cover distribution, and change rapidly with platform movement, making it difficult to accurately acquire and predict channel state information (CSI) in real time.

[0005] Existing channel estimation methods have significant limitations in low-altitude mobile scenarios. Some traditional methods rely on pilot signals for estimation, but with the high-speed movement and frequent relocation of airborne base stations, pilot overhead increases significantly, making it difficult to meet real-time requirements. Other data-driven channel prediction methods (such as deep neural networks and compressed sensing) require training with a large number of measurement samples, but the dynamic nature of the low-altitude environment makes data acquisition costly and limits generalization ability. Furthermore, while radio mapping technology can provide prior information in the spatial domain, traditional interpolation or machine learning methods struggle to construct high-dimensional, high-resolution low-altitude channel distributions, and static maps cannot cope with prediction biases caused by time-varying channels.

[0006] In summary, existing methods struggle to simultaneously achieve low-sample training, time-varying channel modeling, high-dimensional spatial representation, and real-time online calibration capabilities, failing to meet the urgent need for high-precision, low-overhead channel estimation by airborne base stations in low-altitude economic scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide a method for generating radio maps of airborne base stations based on generative artificial intelligence. This method is a novel method for estimating the radio channels of airborne base stations that can construct high-dimensional priors using a small amount of data and perform dynamic inference by combining online observations. This method can improve the reliability and performance of low-altitude mobile communication systems.

[0008] To achieve the above objectives, this invention provides a method for generating radio maps of airborne base stations based on generative artificial intelligence, comprising the following steps: S1, constructing an airborne base station wireless channel model; S2, defining a 9D airborne base station radio map modeling space; S3, using a conditional denoising diffusion probability model to perform offline generative modeling of the 9D airborne base station radio map modeling space; S4, based on the modeling results, constructing an airborne base station radio map without the need for large-scale online measurements, and performing channel estimation for the airborne base station communication system.

[0009] Preferably, S1 specifically refers to: considering the following... AeBSs are 1 air base station A converged air-ground communication system providing radio access services to ground user equipment (UEs) uses a three-dimensional Cartesian coordinate system to model the node positions and motion states: Let the first... airborne base stations 3D position vector at time t With velocity vector They are respectively: ; ;in, They are respectively Time of the first One airborne base station axis, axis, Position coordinates along the axis; They are respectively Time of the first One airborne base station axis, axis, The velocity component along the axial direction; in a discrete-time system, let the time step be... The location of the airborne base station has been updated to: Treating a single terrestrial user equipment (UE) as quasi-static during the radio map building cycle, its three-dimensional position is represented as: ;in, They are respectively Time of the first individual user equipment axis, axis, Position coordinates along the axis The total number of ground user equipment in the system; then the first The first airborne base station and the first The three-dimensional distance between each ground user equipment (UE) is: The probability model for an air-to-ground link being in a line-of-sight (LoS) state is as follows: ;in, The elevation angle between the airborne base station and the ground-based user equipment (UE). , These are empirical parameters related to the propagation environment, used to represent the propagation characteristics of air-to-ground links in different scenarios; the NLoS probability is: Path loss is expressed as: ;in: Due to the movement of the airborne base station, the first The first airborne base station and the first Instantaneous Doppler frequency shift occurs between the UEs: ;in, For carrier frequency, For the speed of light; the channel coefficients are modeled as follows: ;in, Represents the large-scale fading gain, fast fading term The temporal evolution is controlled by the Doppler frequency shift: the LosS scenario is the Rician fading: NLoS scenario refers to Rayleigh decay: ;No. The first air base station to the first When a ground user equipment (UE) transmits a signal, its instantaneous received signal strength (RSS) is used. Represented as: ;in, This refers to the transmission power.

[0010] Preferably, in S2, the 9D airborne base station radio map modeling space is defined to satisfy the following mapping relationship: 9D radio maps are used to describe the mapping relationship of radio channel characteristics between airborne base stations and the spatial distribution of ground users under different spatial deployment and movement states.

[0011] Preferably, S3 specifically involves: explicitly introducing multi-source conditional information into the denoising diffusion probability model to constrain and guide the generation distribution of the radio map, and representing the conditional information as follows: ;in, Indicates the three-dimensional spatial location of the airborne base station; Represents the velocity vector of the airborne base station; Indicates the three-dimensional spatial location of ground user equipment; This indicates the obtained wireless observation information.

[0012] Preferably, in S3, conditional denoising and diffusion probability models are injected with hierarchical embedding and joint coding for different types of conditional information of airborne base stations. Specifically, this involves: for the three-dimensional spatial location of the airborne base station... Velocity vector of airborne base station Three-dimensional spatial location of ground user equipment (UE) First, normalization is performed, and then the result is mapped to a unified high-dimensional feature space using a multilayer perceptron (MLP). A Transformer-based encoder is used for feature extraction, and a self-attention mechanism is used to model the correlation between different observation points to form observation embeddings. The conditional embedding vectors are concatenated or weighted and fused along the feature dimension to obtain a joint conditional vector, which is then injected into the noise prediction network. Multiple noise reduction layers.

[0013] Preferably, in S3, the modeling process of the conditional denoising diffusion probability model includes a forward diffusion process and a backward generation process, specifically: given a real 9D radio map sample The forward process is in Within each time step, a series of latent variables are generated. The transition probability at each step is: ;in, , It is a pre-set noise scheduling parameter used to control the first The noise intensity of the step injection, It is the identity matrix. The mean is Covariance is Gaussian distribution; directly from Sampling yields arbitrary time steps variables ,Right now: ;in, , indicating from the first Step to the first The cumulative noise attenuation coefficient of the step; Consider as raw data and standard Gaussian noise Linear combination form: ;when When large enough, Approaching 0, It approximately follows an isotropic standard Gaussian distribution; the reverse process from Initially, it is generated through a series of progressive denoising operations. Finally, samples that conform to the real data distribution are obtained. The reverse process is modeled as a parameterized Markov chain, and its single-step transition probability is defined as: ;in, These are the parameters of the neural network. Indicates conditional information, It is a learnable denoising mean function. To and The relevant known constants or predefined forms are used; the learning objective of the backward process is reparameterized to predict the noise injected in the forward process. Specifically, the mean of the backward process is expressed as: ;in, This is a noise prediction network; the input is noisy samples. Time step embedding and condition information Output noise The estimate.

[0014] Preferably, in S3, the training objective of the conditional denoising diffusion probability model is to make the reverse generation process approximate the inverse process of the forward diffusion process. Specifically, by minimizing the variational upper bound of the negative log-likelihood, the mean squared error loss function of the conditional denoising diffusion probability model is: ;in, exist{ Uniform sampling in} For training data samples, It is random noise. From the forward formula Calculated.

[0015] Preferably, it also includes an inference step, specifically: based on the trained conditional denoising diffusion probability model, the current three-dimensional spatial position of the airborne base station, the current three-dimensional velocity vector, and the three-dimensional spatial position of the ground user equipment are input, and the corresponding radio map or wireless channel estimation result is output through a multi-step denoising generation method.

[0016] Therefore, the present invention adopts the above-mentioned method for generating radio maps of airborne base stations based on generative artificial intelligence, and the beneficial effects are as follows: (1) The present invention can realize the accurate reconstruction of high-dimensional channel space using a small amount of actual measurement data, significantly reducing the measurement and mapping costs, and improving the feasibility and engineering efficiency of airborne base station deployment.

[0017] (2) By introducing physical state information such as the spatial location and speed of the airborne base station, the present invention constructs a 9D radio map containing the position dimension and speed dimension, realizes the joint modeling of the mobility of the airborne base station and the time-varying characteristics of the channel, overcomes the problem that the existing static radio map is difficult to adapt to high-speed mobile airborne platforms, and improves the accuracy and continuity of channel estimation in dynamic scenarios.

[0018] (3) The present invention generates radio maps of air base stations based on conditional denoising diffusion probability models. In the inference stage, only the current location and speed status of the air base station need to be input to obtain the corresponding channel estimation results, which reduces the dependence on real-time large-scale channel measurement and frequent channel detection, thereby reducing system overhead and improving the real-time adaptability of the air base station communication system under dynamic deployment conditions.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 is an overall flowchart of an embodiment of the method for generating radio maps of airborne base stations based on generative artificial intelligence according to the present invention; Figure 2 is a model training flowchart of an embodiment of the method for generating radio maps of airborne base stations based on generative artificial intelligence according to the present invention; Figure 3 is a reasoning flowchart of an embodiment of the method for generating radio maps of airborne base stations based on generative artificial intelligence according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0023] As shown in Figure 1, a method for generating radio maps of airborne base stations based on generative artificial intelligence includes the following steps: S1, constructing an airborne base station radio channel model, specifically: considering the... Aerial Base Stations (AeBSs) are... A converged air-ground communication system providing radio access services to ground user equipment (UEs) uses a three-dimensional Cartesian coordinate system to model the node positions and motion states: Let the first... airborne base stations 3D position vector at time t With velocity vector They are respectively: ; ;in, They are respectively Time of the first One airborne base station axis, axis, Position coordinates along the axis; They are respectively Time of the first One airborne base station axis, axis, The velocity component in the axial direction.

[0024] In a discrete-time system, let the time step be... The location of the airborne base station has been updated to: .

[0025] Treating a single terrestrial user equipment (UE) as quasi-static during the radio map building cycle, its three-dimensional position is represented as follows: ;in, They are respectively Time of the first individual user equipment axis, axis, Position coordinates along the axis This represents the total number of ground user equipment in the system.

[0026] Then the first The first airborne base station and the first The three-dimensional distance between each ground user equipment (UE) is: .

[0027] The probability model for the air-to-ground link being in a line-of-sight (LoS) state is as follows: ;in, The elevation angle between the airborne base station and the ground-based user equipment (UE). , These are empirical parameters related to the propagation environment, i.e., environmental parameters, used to represent the propagation characteristics of air-to-ground links in different scenarios; their specific values ​​depend on the type of environment considered. The parameter values ​​corresponding to various typical environments are shown in Table 1 below.

[0028] Table 1 Environmental Parameter Values

[0029] The probability of NLoS is: .

[0030] Path loss is expressed as: ;in: .

[0031] Due to the movement of the airborne base station, the first The first airborne base station and the first Instantaneous Doppler frequency shift occurs between the UEs: ;in, For carrier frequency, It is the speed of light.

[0032] The channel coefficients are modeled as follows: ;in, Represents the large-scale fading gain, fast fading term The temporal evolution is controlled by the Doppler frequency shift: the LosS scenario is the Rician fading: .

[0033] NLoS scenario, also known as Rayleigh decay: .

[0034] No. The first air base station to the first When a ground user equipment (UE) transmits a signal, its instantaneous received signal strength (RSS) is used. Represented as: ;in, This refers to the transmission power.

[0035] S2. Define the 9D aerial base station radio map modeling space to satisfy the following mapping relationship: This 9D radio map is used to describe the mapping relationship between the wireless channel characteristics of airborne base stations and the spatial distribution of ground users under different spatial deployment and movement states, thereby providing support for coverage optimization, interference management and resource allocation of mobile airborne base station communication systems.

[0036] S3. In the context of airborne base station communication, the characteristics of the wireless channel are not only affected by spatial geometry but also closely related to the motion state and temporal evolution of the airborne base station. In particular, the flight speed of the airborne base station significantly alters the Doppler shift, channel coherence time, and fast fading statistical distribution, resulting in significant differences in the radio map under different motion states. Therefore, this invention employs a conditional denoising diffusion probability model to perform offline generative modeling of the 9D airborne base station radio map modeling space. Specifically, multi-source conditional information is explicitly introduced into the denoising diffusion probability model to constrain and guide the generation distribution of the radio map, thereby achieving refined modeling of the dynamic airborne communication environment. This invention represents the conditional information as follows: ;in, Indicates the three-dimensional spatial location of the airborne base station; Represents the velocity vector of the airborne base station; Indicates the three-dimensional spatial location of ground user equipment; This represents the obtained wireless observation information, such as sparse RSS, SINR, or other link quality measurement segments, used to provide prior constraints for radio map generation.

[0037] At the model implementation level, this invention employs a hierarchical embedding and joint coding approach to inject conditional denoising and diffusion probability model for different types of conditional information of airborne base stations. Specifically, this involves: injecting continuous values ​​of airborne base stations, low-dimensional physical state conditions, and three-dimensional spatial location into the model. Velocity vector of airborne base station Three-dimensional spatial location of ground user equipment (UE) First, normalization is performed, and then the data is mapped to a unified high-dimensional feature space through a multilayer perceptron (MLP) to characterize the geometric relationship between the airborne base station and the user, their relative motion characteristics, and their nonlinear effects on path loss, Doppler shift, and channel time correlation.

[0038] Wireless observation information Considering its sparsity and irregularity in space and time, a Transformer-based encoder is used for feature extraction, and a self-attention mechanism is used to model the correlation between different observation points to form an observation embedding representation with global context information.

[0039] Subsequently, the conditional embedding vectors are concatenated or weighted and fused along the feature dimension to obtain a joint conditional vector, which is then injected into the noise prediction network. The model employs multiple denoising layers to modulate the generated distribution throughout the backdiffusion process. Through the aforementioned conditional information modeling and embedding mechanism, the model can explicitly perceive the spatial location, motion state, and prior observations of airborne base stations, enabling high-fidelity construction of 9D radio maps in dynamic airborne communication environments.

[0040] The conditional denoising diffusion probability model, as a deep generative model, learns the data distribution by simulating a progressive noise addition and removal process. Its framework includes two key processes: a fixed forward diffusion process and a learnable backward generation process. Specifically, given real 9D radio map samples... The forward process is in Within each time step, a series of latent variables are generated. The transition probability at each step is: ;in, , It is a pre-set noise scheduling parameter used to control the first The noise intensity of the step injection. It is the identity matrix. The mean is Covariance is The Gaussian distribution.

[0041] Because the Gaussian distribution has good closure properties in this process, it can be directly obtained from... without step-by-step iteration. Sampling yields arbitrary time steps variables ,Right now: ;in, , indicating from the first Step to the first The cumulative noise attenuation coefficient of the step; equivalently, will Consider as raw data and standard Gaussian noise Linear combination form: ;when When large enough, Approaching 0, It approximately follows an isotropic standard Gaussian distribution.

[0042] Denoising process: Reverse process from Initially, it is generated through a series of progressive denoising operations. Finally, samples that conform to the real data distribution are obtained. This reverse process is modeled as a parameterized Markov chain, and its single-step transition probability is defined as: ;in, These are the parameters of the neural network. Indicates conditional information, It is a learnable denoising mean function. It is usually set to be with The relevant known constants or predefined forms.

[0043] The key idea of ​​the conditional denoising diffusion probability model lies in reparameterizing the learning objective of the backward process as predicting the noise injected in the forward process. Specifically, the mean of the backward process can be expressed as: ;in, This is a noise prediction network, which typically uses a U-Net structure and takes noisy samples as input. Time step embedding and condition information Output noise The estimate.

[0044] As shown in Figure 2, the training objective of the conditional denoising diffusion probability model is to make the reverse generation process approximate the inverse of the forward diffusion process as closely as possible. Specifically, by minimizing the variational upper bound of the negative log-likelihood, the mean squared error loss function of the conditional denoising diffusion probability model is: ;in, exist{ Uniform sampling in} For training data samples, It is random noise. From the forward formula Calculated.

[0045] In each iteration, the algorithm randomly selects a time step. Calculate the data after adding noise. Then let the noise prediction network The model predicts the added noise. By continuously minimizing the difference between the predicted noise and the actual noise, the model gradually learns the denoising rules under different noise intensities, thereby achieving high-quality conditional generation.

[0046] As shown in Figure 2, during the training process, the three-dimensional spatial location information, three-dimensional velocity information of the airborne base station, and the three-dimensional spatial location information of the user equipment are conditionally embedded and encoded. The embedded conditional features are then input into the conditional denoising diffusion probability model. By gradually adding noise to the wireless measurement data and learning the corresponding denoising process, the conditional denoising diffusion probability model learns the statistical distribution relationship between the 9D modeling space and the wireless channel observation information at different noise levels. This guides the generation process of the radio map to conform to the spatial deployment status and motion characteristics of the airborne base station.

[0047] As shown in Figure 3, during the inference phase, based on the trained conditional denoising diffusion probability model, the current three-dimensional spatial position of the airborne base station, the current three-dimensional velocity vector, and the three-dimensional spatial position of the ground user equipment are input. The corresponding radio map or wireless channel estimation result is output through a multi-step denoising generation method to achieve rapid perception of the communication environment of the airborne base station.

[0048] S4. Based on the modeling results, by learning from a small amount of air-to-ground link measurement data, a mapping relationship is established between the deployment status and motion characteristics of airborne base stations and the distribution of wireless channels. This enables the construction of an airborne radio map without the need for large-scale online measurements and provides stable prior information for channel estimation.

[0049] Therefore, the present invention adopts the above-mentioned method for generating radio maps of airborne base stations based on generative artificial intelligence. Compared with existing channel estimation methods based on interpolation, regression or discriminative learning, the present invention can obtain high-precision spatiotemporal channel estimation with low measurement overhead. It can effectively support channel estimation, coverage optimization and resource allocation in airborne base station communication systems and is applicable to communication systems of airborne base stations and other high-speed mobile wireless platforms.

[0050] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating radio maps of airborne base stations based on generative artificial intelligence, characterized in that, Includes the following steps: S1. Construct the wireless channel model of the airborne base station; S2. Define the 9D airborne base station radio map modeling space; S3. The conditional denoising diffusion probability model is used to perform offline generative modeling of the 9D airborne base station radio map modeling space. S4. Based on the modeling results, construct an aerial base station radio map without the need for large-scale online measurements, and perform channel estimation for the aerial base station communication system.

2. The method for generating radio maps of airborne base stations based on generative artificial intelligence according to claim 1, characterized in that, S1 specifically refers to: considering the... AeBSs are 1 air base station A converged air-ground communication system providing radio access services to ground user equipment (UEs) uses a three-dimensional Cartesian coordinate system to model the node positions and motion states: Let the first... airborne base stations 3D position vector at time t With velocity vector They are respectively: ； ;in, They are respectively Time of the first One airborne base station axis, axis, Position coordinates along the axis; They are respectively Time of the first One airborne base station axis, axis, The velocity component along the axial direction; in a discrete-time system, let the time step be... The location of the airborne base station has been updated to: Treating a single terrestrial user equipment (UE) as quasi-static during the radio map building cycle, its three-dimensional position is represented as: ;in, They are respectively Time of the first individual user equipment axis, axis, Position coordinates along the axis The total number of ground user equipment in the system; then the first The first airborne base station and the first The three-dimensional distance between each ground user equipment (UE) is: The probability model for an air-to-ground link being in a line-of-sight (LoS) state is as follows: ;in, The elevation angle between the airborne base station and the ground-based user equipment (UE). 、 These are empirical parameters related to the propagation environment, used to represent the propagation characteristics of air-to-ground links in different scenarios; the NLoS probability is: Path loss is expressed as: ;in: Due to the movement of the airborne base station, the first The first airborne base station and the first Instantaneous Doppler frequency shift occurs between the UEs: ;in, For carrier frequency, For the speed of light; the channel coefficients are modeled as follows: ;in, Represents the large-scale fading gain, fast fading term The temporal evolution is controlled by the Doppler frequency shift: the LosS scenario is the Rician fading: NLoS scenario refers to Rayleigh decay: ; the The first air base station to the first When a ground user equipment (UE) transmits a signal, its instantaneous received signal strength (RSS) is used. Represented as: ;in, This refers to the transmission power.

3. The method for generating radio maps of airborne base stations based on generative artificial intelligence according to claim 2, characterized in that, In S2, the 9D airborne base station radio map modeling space is defined to satisfy the following mapping relationship: 9D radio maps are used to describe the mapping relationship of radio channel characteristics between airborne base stations and the spatial distribution of ground users under different spatial deployment and movement states.

4. The method for generating radio maps of airborne base stations based on generative artificial intelligence according to claim 3, characterized in that, S3 specifically involves: explicitly introducing multi-source conditional information into the denoising diffusion probability model to constrain and guide the generation distribution of the radio map, and representing the conditional information as follows: ;in, Indicates the three-dimensional spatial location of the airborne base station; Represents the velocity vector of the airborne base station; Indicates the three-dimensional spatial location of ground user equipment; This indicates the obtained wireless observation information.

5. The method for generating radio maps of airborne base stations based on generative artificial intelligence according to claim 4, characterized in that, In S3, a hierarchical embedding and joint coding approach is used to inject conditional denoising and diffusion probability model for different types of conditional information of airborne base stations. Specifically, this involves: the three-dimensional spatial location of the airborne base station... Velocity vector of airborne base station Three-dimensional spatial location of ground user equipment (UE) First, normalization is performed, and then the result is mapped to a unified high-dimensional feature space through a multilayer perceptron (MLP). right A Transformer-based encoder is used for feature extraction, and a self-attention mechanism is used to model the correlation between different observation points to form observation embeddings. The conditional embedding vectors are concatenated or weighted and fused along the feature dimension to obtain a joint conditional vector, which is then injected into the noise prediction network. Multiple noise reduction layers.

6. The method for generating an aerial base station radio map based on generative artificial intelligence according to claim 5, characterized in that, In S3, the modeling process of the conditional denoising diffusion probability model includes a forward diffusion process and a backward generation process, specifically: given a real 9D radio map sample The forward process is in Within each time step, a series of latent variables are generated. The transition probability at each step is: ;in, ， It is a pre-set noise scheduling parameter used to control the first The noise intensity of the step injection, It is the identity matrix. The mean is Covariance is Gaussian distribution; directly from Sampling yields arbitrary time steps variables ,Right now: ;in, , indicating from the first Step to the first The cumulative noise attenuation coefficient of the step; Consider as raw data and standard Gaussian noise Linear combination form: ;when When large enough, Approaching 0, It approximately follows an isotropic standard Gaussian distribution; the reverse process from Initially, it is generated through a series of progressive denoising operations. Finally, samples that conform to the real data distribution are obtained. The reverse process is modeled as a parameterized Markov chain, and its single-step transition probability is defined as: ;in, These are the parameters of the neural network. Indicates conditional information, It is a learnable denoising mean function. To and The relevant known constants or predefined forms are used; the learning objective of the backward process is reparameterized to predict the noise injected in the forward process. Specifically, the mean of the backward process is expressed as: ;in, This is a noise prediction network; the input is noisy samples. Time step embedding and condition information Output noise The estimate.

7. The method for generating an aerial base station radio map based on generative artificial intelligence according to claim 6, characterized in that, In S3, the training objective of the conditional denoising diffusion probability model is to make the reverse generation process approximate the inverse of the forward diffusion process. Specifically, by minimizing the variational upper bound of the negative log-likelihood, the mean squared error loss function of the conditional denoising diffusion probability model is: ;in, exist{ Uniform sampling in} For training data samples, It is random noise. From the forward formula Calculated.

8. The method for generating radio maps of airborne base stations based on generative artificial intelligence according to claim 7, characterized in that, It also includes an inference step, specifically: based on the trained conditional denoising diffusion probability model, the current three-dimensional spatial position of the air base station, the current three-dimensional velocity vector, and the three-dimensional spatial position of the ground user equipment are input, and the corresponding radio map or wireless channel estimation result is output through a multi-step denoising generation method.

Images