Unmanned aerial vehicle-to-ground channel prediction method and system based on adaptive fourier features

Abstract

The application relates to the technical field of channel modeling and prediction, and provides a UAV (Unmanned Aerial Vehicle) ground channel prediction method and system based on adaptive Fourier features, which comprises the following steps: acquiring a historical channel impulse response sequence, extracting a timestamp and a multipath time delay, respectively performing adaptive Fourier feature mapping, and identifying context information of the time delay of different path signals in multipath propagation; modeling the correlation between different multipath signals based on the time delay AFF mapping, constructing a cross-path kernel to mix the historical channel impulse response sequence, and extracting time context coding; fusing the time context coding, the time AFF mapping and the context information of the time delay, and performing regression prediction to obtain a channel state at a future moment. Through the introduction of the adaptive Fourier feature mapping, the time delay semantic vector and the cross-path kernel mechanism, and the fusion of the GRU time sequence coding and the multi-feature fusion strategy, the application realizes accurate modeling and efficient prediction of the fast time-varying characteristics of the UAV ground channel.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of channel modeling and prediction, in particular to a UAV-to-ground channel prediction method and system based on adaptive Fourier features. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] With the large-scale deployment of UAV technology in the fields of civil, industrial and public safety, air-to-ground wireless communication has become the core infrastructure to ensure the safe flight and task execution of UAVs. When UAVs complete tasks such as aerial inspection, logistics transportation, disaster assessment and communication relay, they must rely on stable and reliable wireless links to implement remote control instruction issuance, flight state monitoring and business data backhaul. However, compared with traditional ground mobile communication, UAVs have the characteristics of high flight altitude, fast moving speed and frequent environmental changes, which make the air-to-ground signal in the propagation process be affected by path loss, multipath reflection, ground obstruction and Doppler shift, etc. Multiple factors, resulting in a highly dynamic, strong random and strong time-varying channel.

[0004] In complex airspace environments, the communication reliability of UAVs depends not only on the current channel quality, but also on the prediction ability of the communication system for future channel changes. Accurate channel modeling and real-time channel estimation can provide the communication system with a description and prediction of the wireless propagation environment, so that the system can know the trend of link quality changes in advance, and thus take more robust communication strategies. For example, by predicting that the UAV may enter a weak coverage area through the channel model, the system can adjust the transmission power, switch the link mode or adjust the modulation and coding mode in advance. In addition, channel modeling plays a key role in communication resource allocation. Due to the increasing number of UAVs and the diversification of business needs, the ground control end or cellular base station needs to allocate spectrum, time slots or power resources to different UAVs according to the channel prediction results, in order to improve the overall network throughput and link stability. Therefore, the channel model is not only a tool to describe the physical propagation characteristics, but also an important basis for the intelligent scheduling and resource optimization of the communication system, which realizes the core logic chain of "physical channel modeling → wireless environment prediction → resource allocation strategy optimization", and is directly related to the reliability, real-time performance and spectrum utilization efficiency of the UAV communication network.

[0005] Although the rapid development of wireless communication and artificial intelligence technology in recent years provides new means for unmanned aerial vehicle channel modeling, the existing channel modeling and prediction research still has many deficiencies. In terms of modeling objects, most of the current channel research focuses on ground cellular communication scenarios, and the systematic research on unmanned aerial vehicle air-to-ground channel is relatively limited. In terms of channel characteristics, the received signal power during the flight of the unmanned aerial vehicle is affected by multipath propagation and fading in the delay domain, and is affected by the Doppler effect in the time domain, showing strong randomness and rapid change characteristics, and the channel impulse response contains a large number of high-frequency components in the time axis and the multipath delay axis. The traditional prediction method based on neural network is limited by the inherent spectral bias problem, and it is more likely to learn low-frequency change characteristics, and it is difficult to effectively depict the high-frequency time-varying behavior in the unmanned aerial vehicle channel, resulting in insufficient modeling and prediction accuracy of the fast time-varying channel. SUMMARY

[0006] In order to solve the above problems, the present application provides a kind of unmanned aerial vehicle channel prediction method and system based on adaptive fourier feature, by introducing adaptive fourier feature mapping, delay semantic vector and cross-path kernel mechanism, fusion GRU time sequence coding and multi-feature fusion strategy, realize the accurate modeling and efficient prediction of the fast time-varying characteristics of unmanned aerial vehicle channel.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0008] The first aspect of the present application provides a kind of unmanned aerial vehicle channel prediction method based on adaptive fourier feature, including the following steps:

[0009] The unmanned aerial vehicle channel to be predicted is obtained, and the measured historical channel impulse response sequence is obtained;

[0010] The timestamp corresponding to each historical channel impulse response and the multipath delay are extracted, and adaptive fourier feature mapping is carried out respectively to obtain time AFF mapping and delay AFF mapping;

[0011] The channel impulse response of the last frame channel signal and the delay AFF mapping are used to define the delay semantic vector, and the delay context information of different path signals in multipath propagation is extracted;

[0012] The correlation between different multipath signals is modeled based on the delay AFF mapping, the cross-path kernel is constructed to mix the historical channel impulse response sequence, and the new cross-path channel representation is input into the GRU time sequence encoder to extract the time context code;

[0013] The time context code, time AFF mapping and context information of delay are fused to obtain the fused spatiotemporal features, and the channel state at future time is obtained by regression prediction.

[0014] The second aspect of the present application provides a UAV-to-ground channel prediction system based on adaptive Fourier features, comprising:

[0015] A measurement data acquisition module is configured to acquire a UAV-to-ground channel to be predicted and acquire a measured historical channel impulse response sequence;

[0016] An AFF mapping module is configured to extract a timestamp and a multipath delay corresponding to each historical channel impulse response, respectively perform adaptive Fourier feature mapping, and obtain a time AFF mapping and a delay AFF mapping;

[0017] A context information extraction module is configured to define a delay semantic vector based on a channel impulse response of a previous frame of channel signals and the delay AFF mapping, and extract time delay context information of different path signals in multipath propagation;

[0018] A time sequence encoding module is configured to model a correlation between different multipath signals based on the delay AFF mapping, construct a cross-path kernel for mixing the historical channel impulse response sequence, obtain a new channel representation after cross-path, input the new channel representation into a GRU time sequence encoder, and extract time context encoding;

[0019] A fusion and regression prediction module is configured to fuse the time context encoding, the time AFF mapping, and the context information of the delay, obtain fused spatio-temporal features, and perform regression prediction to obtain a channel state at a future time.

[0020] The third aspect of the present application provides a UAV-to-ground channel prediction system based on adaptive Fourier features, comprising a collection device and a processor.

[0021] The collection device is used to collect a historical channel impulse response sequence.

[0022] The processor is configured to perform the steps of the UAV-to-ground channel prediction method based on adaptive Fourier features described above.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] The method of the present application can effectively deal with the high-frequency oscillation characteristics of the time domain and the delay domain in the unmanned aerial vehicle channel by using the AFF mapping, improve the perception ability of the model to the fast time-varying channel, and solve the problem that the traditional model is difficult to capture high-frequency changes; by constructing the delay semantic vector and the span kernel function, the modeling ability of the model to the multipath structure change is enhanced, the semantic dependence and dynamic correlation between paths can be more comprehensively captured, and the limitation of the existing model to the dynamic modeling of multipath is overcome; the GRU encoder is introduced to capture the time context information, further improve the timing consistency and accuracy of the prediction, effectively improve the continuity under multi-time prediction; the multi-feature fusion improves the spatio-temporal feature representation ability, so that the model has stronger generalization and robustness in complex scenes, thereby being applicable to various environment and task scenes, and relieving the limitation of the existing method depending on specific scenes.

[0025] The advantages of the present application and the advantages of the additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0026] The drawings accompanying the specification of the present application form a part thereof and serve to provide further understanding of the present application, the illustrative embodiments of the present application and its description serve to explain the present application, and do not constitute limitations thereof.

[0027] Figure 1 is a flow chart of the unmanned aerial vehicle channel prediction method based on adaptive Fourier features of embodiment 1 of the present application;

[0028] Figure 2 is a flow chart of the unmanned aerial vehicle channel prediction method based on adaptive Fourier features of embodiment 1 of the present application;

[0029] Figure 3 is a structure diagram of the unmanned aerial vehicle channel data enhancement module based on the generative adversarial network of embodiment 1 of the present application;

[0030] Figure 4 is a comparison diagram of the channel root mean square delay spread generated under different scenarios in the simulation example of embodiment 1 of the present application and the real channel root mean square delay spread;

[0031] Figure 5 is a comparison diagram of the channel Rician factor generated under different scenarios in the simulation example of embodiment 1 of the present application and the real channel Rician factor. DETAILED DESCRIPTION

[0032] The present application will be further described below in combination with the drawings and embodiments.

[0033] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. It will be apparent that systems and / or processes described herein can be implemented in various forms of hardware, software, firmware, special purpose processors, or a combination of these, and that the methods described herein can be implemented by one or more computer programs executing on one or more computers or other programmable devices. The terms "first", "second", "third", "fourth", "fifth", "sixth", etc. do not necessarily mean that the corresponding clause is the first, second, third, fourth, fifth, sixth, etc. in time or space, but are only identifiers differentiating between different features. The embodiments of the present application can be combined with each other, if not in conflict.

[0035] Technical term explanation:

[0036] Channel: refers to the comprehensive performance of the entire physical process and propagation environment experienced by the wireless signal in the propagation process, which is used to describe the amplitude attenuation, phase rotation, frequency offset, time delay spread, and random fading effects of electromagnetic waves from the transmitting end to the receiving end. The channel model uses a model to describe the above propagation characteristics, which is the basis for realizing the performance analysis, link prediction and parameter optimization of the communication system.

[0037] Spectral bias: the phenomenon that neural networks tend to preferentially fit low-frequency components while having relatively insufficient fitting ability for high-frequency components when learning function mapping. In the channel prediction task, spectral bias will cause the model to more easily learn the smooth and slowly changing trends in channel changes, while it is difficult to accurately depict the rapid fluctuations and high-frequency variation characteristics caused by high-speed motion, Doppler effect and complex multipath propagation.

[0038] Time domain: refers to the dimension of describing channel characteristics with time as the independent variable, which is used to characterize the dynamic characteristics of wireless channels evolving over time. In the unmanned aerial vehicle communication scenario, the time domain mainly reflects the rapid change behavior of the channel over time due to factors such as unmanned aerial vehicle motion, Doppler frequency shift and environmental changes.

[0039] Delay domain: refers to the dimension of describing channel characteristics with signal propagation delay as the independent variable, which is used to characterize the multipath propagation structure and relative delay characteristics of each propagation path in the wireless channel. In the unmanned aerial vehicle air-to-ground channel, the delay domain reflects the difference in arrival time of different multipath components and their energy distribution, which is an important basis for describing multipath fading and power delay spectrum.

[0040] Embodiment 1

[0041] In the technical solutions disclosed in one or more embodiments, such asFigures 1 to 5 As shown in the figure, a UAV-to-ground channel prediction method based on adaptive Fourier features includes the following steps:

[0042] Step 1, obtaining the UAV-to-ground channel to be predicted, and obtaining the measured historical channel impulse response sequence;

[0043] Step 2, extracting the timestamp corresponding to each historical channel impulse response And the multipath time delay , respectively, adaptive Fourier feature mapping (abbreviated as AFF mapping) is carried out, time AFF mapping And time delay AFF mapping are obtained;

[0044] Step 3, defining the time delay semantic vector based on the channel impulse response of the last frame of channel signal and the time delay AFF mapping, and extracting the context information of the time delay of different path signals in multipath propagation ;

[0045] Step 4, modeling the correlation between different multipath signals based on the time delay AFF mapping, constructing the cross-path kernel to mix the historical channel impulse response sequence, and obtaining the new channel representation after cross-path Input into the GRU (Gated Recurrent Unit) time sequence encoder to extract the time context code ;

[0046] Step 5, fusing the time context code , time AFF mapping And the context information of the time delay Obtain the fused space-time features, and perform regression prediction to obtain the channel state at the future time;

[0047] In this embodiment, the time-varying characteristics of the UAV-to-ground wireless channel are modeled based on adaptive Fourier features (AFF). First, the time stamp and time delay parameters are extracted from the measured historical channel impulse response. In view of the high-frequency oscillation characteristics in the time domain and the time delay domain, the time stamp t and the time delay τ are respectively subjected to adaptive Fourier feature mapping, so as to obtain frequency domain features with better representation ability. Then, the last frame of channel state and the time delay mapping are used to construct a time delay semantic vector, which explicitly captures the time structure relationship of different paths in the multipath signal. At the same time, in order to model the correlation between different paths, a cross-path kernel function is constructed to nonlinearly mix multiple radial signals, forming a new channel representation. This representation is input into the GRU time sequence encoder to further extract the dynamic context features in the time sequence. Finally, the fused features are jointly input into the regression network to predict the channel state at the future time, thereby realizing accurate modeling of the high-speed dynamic channel.

[0048] The embodiment can effectively cope with the high-frequency oscillation characteristics of the time domain and the time delay domain in the unmanned aerial vehicle channel by using the AFF mapping, improve the perception ability of the model to the fast time-varying channel, and solve the problem that the traditional model is difficult to capture high-frequency changes; by constructing the time delay semantic vector and the span kernel function, the modeling ability of the model to the multi-path structure change is enhanced, the semantic dependence and dynamic correlation between paths can be more comprehensively captured, and the limitation of the existing model to the multi-path dynamic modeling is overcome; the GRU encoder is introduced to capture the time context information, which further improves the timing consistency and accuracy of the prediction, effectively improves the continuity under multi-time prediction; the multi-feature fusion improves the spatio-temporal feature representation ability, so that the model has stronger generalization and robustness in complex scenes, thereby being applicable to various environment and task scenes, and relieving the limitation of the existing method on specific scenes.

[0049] In view of the rapid change characteristics of the channel characteristics in the time domain and the time delay domain, the embodiment introduces adaptive Fourier feature mapping in the time and time delay dimensions, maps the time and time delay parameters to a high-dimensional, learnable frequency domain feature space, effectively alleviates the problem of insufficient modeling ability of the traditional neural network to high-frequency components, improves the representation ability of the model to the rapid change caused by the Doppler effect and multipath fading, and thus realizes accurate modeling and prediction of the rapid time-varying characteristics of the unmanned aerial vehicle air-to-ground channel.

[0050] In the above steps, the channel prediction model fusing time-delay AFF and GRU is also implemented, and further includes constructing a channel prediction model fusing time-delay AFF and GRU, including: an adaptive Fourier transform module, a GRU (Gated Recurrent Unit) time sequence encoder, and a multilayer perception regressor.

[0051] The adaptive Fourier transform module is configured to perform adaptive Fourier feature mapping (AFF mapping for short) on the time stamp t and the multipath time delay τ corresponding to each historical impulse response, respectively, to obtain time AFF mapping and time delay AFF mapping ; the time delay semantic vector is defined based on the last frame of channel signals, and the context information of the time delay of different path signals in multipath propagation is extracted ; the correlation between different multipath signals is modeled based on the time delay AFF mapping, a span kernel is constructed to mix the historical channel impulse response sequence, and a new channel representation after span is obtained .

[0052] The GRU (Gated Recurrent Unit) time sequence encoder is used to receive the channel representation after span mixed by the span, and extract the context code of time.

[0053] a multi-layer perceptron regressor is used to regress the fused spatio-temporal features to predict the channel state at the future time, such as the channel impulse response;

[0054] Further, an adversarial network is further included to perform data augmentation on the obtained training data.

[0055] In step 1, the acquisition of the historical channel impulse response sequence of the embodiment is performed by a channel measurement system including an air signal transmitting unit and a ground receiving unit. In actual use, for the unmanned aerial vehicle communication scenario, the signal transmitting end of the unmanned aerial vehicle is the air signal transmitting unit.

[0056] The transmitting unit and the receiving unit are respectively software defined radio devices, and are uniformly configured and cooperatively controlled through a GNU Radio platform. The system transmitting signal is a pseudo-random sequence with good autocorrelation characteristics, which is transmitted by the unmanned aerial vehicle after being up-converted to the carrier frequency. The ground receiving unit down-converts and baseband processes the signal, and then extracts the channel response through correlation operation. By setting the transmitting and receiving antenna gain, system sampling rate, unmanned aerial vehicle flight height, initial horizontal distance, flight trajectory, etc., multi-band unmanned aerial vehicle channel measurement is carried out.

[0057] In step 1, the received signal and the transmitting signal are estimated to obtain the channel impulse response , the process is as follows:

[0058] Step 11, the received signal is represented by the transmitting signal and the channel impulse response , the formula is:

[0059] (1);

[0060] wherein, denotes the time delay, and t denotes the time;

[0061] Step 12, the transmitting signal is taken as white noise, and the autocorrelation of the transmitting signal is calculated, the function is:

[0062] (2);

[0063] wherein, is the white noise power density, is the Dirac function, is the autocorrelation function of the transmitting signal ;

[0064] Step 13, based on the autocorrelation of the transmitting signal, the received signal and the transmitting signal The cross-correlation is expressed as:

[0065] (3) ;

[0066] wherein, denotes a convolution operation, , denote the integral variable and the time delay, respectively.

[0067] Substituting equation (2) into equation (3), based on the assumption that the transmitted signal is white noise, the autocorrelation function thereof is an impulse function, after being substituted into the cross-correlation expression, the sampling property of the Dirac function is used to obtain that the cross-correlation function of the received signal and the transmitted signal is proportional to the channel impulse response, thereby realizing the estimation of the channel impulse response, and equation (4) is obtained:

[0068] (4) ;

[0069] Step 14, based on the cross-correlation, the channel impulse response is obtained, and the channel impulse response can be estimated as:

[0070] (5) ;

[0071] The method of the embodiment utilizes the good correlation property of the pseudo-random sequence, and realizes the effective estimation of the channel impulse response in a complex air-to-ground environment.

[0072] In step 2, the AFF in the time domain and the time delay domain is introduced to capture the timing change and the frequency selective fading in the wireless channel.

[0073] In step 2, the time stamp corresponding to each historical channel impulse response and the multipath time delay are extracted, and adaptive Fourier feature mapping (AFF mapping for short) is respectively performed to obtain the time AFF mapping and the time delay AFF mapping The method comprises the following steps:

[0074] Step 21, time AFF mapping: the time stamp corresponding to the channel impulse response is subjected to adaptive Fourier feature mapping to obtain the timing dynamic change information of the channel, i.e., the time AFF mapping .

[0075] For the normalized time , the time AFF mapping is defined as:

[0076] (6) ;

[0077] wherein, is the base of the time AFF, These represent the learnable frequency and phase of the time-dependent AFF, respectively.

[0078] Step 22: Using adaptive Fourier transform, the time delay of the channel impulse response of different paths in multipath propagation is mapped to obtain the time delay AFF mapping;

[0079] For each normalized delay The delay AFF mapping is defined as follows:

[0080] (7);

[0081] in, The base of the delay AFF is represented. These represent the learnable frequency and phase of the delay AFF, respectively;

[0082] Normalized delay refers to the result of normalizing the delay values ​​of signals from different paths in multipath propagation. It is used to characterize the relative delay differences of each path and is key data for capturing the spatial characteristics of multipath.

[0083] Step 3, the method for extracting the contextual information of the time delay of signals from different paths in multipath propagation, includes:

[0084] Step 31: Stack the delay AFF mapping features of all multipaths into a matrix to obtain the delay AFF dictionary, represented as:

[0085] (8);

[0086] in, Indicates transpose. It is the m-th normalized time delay grid point. , D It is the number of multipath paths. Delay AFF dictionary. It is a matrix;

[0087] Each dimension of the channel impulse response corresponds to a multipath component, and each multipath has its own time delay. To model the structural characteristics between multipaths, a time delay feature representation is generated for each multipath to improve the comprehensiveness of the modeling features.

[0088] Step 32: Channel impulse response and delay AFF dictionary based on the channel signal of the previous frame. Define a time delay semantic vector to obtain contextual information about the time delay of signals from different paths in multipath propagation. ;

[0089] (9);

[0090] in, This represents the transpose of the delay AFF dictionary. channel impulse response of the previous frame; base of the delay AFF;

[0091] In this step, the channel impulse response of the previous frame is projected to provide prior information about the delay, helping the model to conduct timing modeling.

[0092] In step 4, the correlation between different multipath signals is modeled based on the delay AFF mapping, and a cross-path kernel is constructed to mix the historical channel impulse response sequence, obtaining a new channel representation after cross-path , comprising the following steps:

[0093] Step 41, based on the delay AFF dictionary obtained in step 31, a learnable cross-path kernel is defined on the delay axis based on the delay AFF dictionary to model the correlation between different paths, represented as follows:

[0094] (10);

[0095] where A is a positive semi-definite matrix representing the structure of the cross-path kernel, which can be selected as a diagonal form to reduce computational complexity, represents the cross-path kernel, represents the delay AFF dictionary, represents the transpose, D is the number of multipaths; through the cross-path kernel , the model can learn the correlation between different multipaths, thereby enhancing the robustness to multipath interference.

[0096] Each sample point of the channel impulse response corresponds to the delay of a different propagation path in the wireless channel; arranging these delays in chronological order is the delay axis.

[0097] Step 42, for any historical channel impulse response , by multiplying with the cross-path kernel, a new channel representation is obtained, i.e., the historical channel response representation sequence after cross-path:

[0098] (11);

[0099] where L is the length of the history window, representing the time series window size used for prediction.

[0100] The above implementation of the embodiment can explicitly model the structured correlation between multiple paths by introducing a delay AFF mapping and a learnable span kernel on the delay axis, thereby significantly improving the accuracy and robustness of channel impulse response prediction. First, the delay AFF maps the scalar delay to a continuous high-dimensional learnable feature space, so that the multiple path delays are distributed in the feature domain to form a capturable geometric structure; then, the span kernel constructed by the delay AFF dictionary matrix and the kernel matrix can effectively describe the degree of correlation between different multiple path paths, and realize the weighted mixing of multiple path gains. This process is equivalent to introducing a multiple path structure prior into the historical channel sequence, so that the channel representation after span kernel processing has both physical meaning and stronger separability.

[0101] In step 4, further, the channel representation after span is input into a GRU (Gated Recurrent Unit) time sequence encoder to extract a time context code , and the formula is:

[0102] (12);

[0103] wherein, represents a gated recurrent unit for extracting a time context code, represents a parameter set of the GRU;

[0104] In step 5, the time context code, the time AFF mapping, and the delay semantic vector are fused to obtain a fused space-time feature, and the formula is:

[0105] (13);

[0106] The fused space-time feature is input into a multilayer perceptron regressor to obtain a predicted value of the channel impulse response at the next time:

[0107] (14);

[0108] Further, the channel prediction model fusing the time-delay AFF and the GRU is fused, and the training target of the model is to minimize the combination of the weighted L2 loss and the SmoothL1 loss, and the total loss function is represented as follows:

[0109] (15);

[0110] wherein, is a weight factor for balancing the two loss functions.

[0111] Further, the process of training the channel prediction model fusing the time-delay AFF and the GRU further includes the following:

[0112] Step S1, obtaining a to-be-predicted unmanned aerial vehicle-to-ground channel, and obtaining a measured historical channel impulse response sequence;

[0113] Step S2, constructing an adversarial learning network as a generation model, performing data enhancement on the obtained historical channel impulse response sequence data, and constructing a data set for training;

[0114] Step S3, based on the enhanced data set, taking channel impulse response data obtained in a previous time period as input and taking channel impulse response at a target time in a subsequent time as output, training the constructed channel prediction model of the fusion of time-delay AFF and GRU, and obtaining the trained channel prediction model of the fusion of time-delay AFF and GRU;

[0115] The training method of the embodiment effectively solves the problem of insufficient training data in deep learning channel prediction by using the adversarial learning network to perform data enhancement on the historical channel impulse response sequence. Traditional methods usually rely on data generated by simulators or ray tracing for training, which has controllability and repeatability, but there is still a significant distribution difference between the real air-to-ground scene, which limits the generalization ability of the model. Real channel measurement is also restricted by factors such as expensive equipment, high measurement cost, and complex deployment, making it difficult to obtain a sufficient number of training samples with rich diversity. The embodiment introduces an adversarial learning network, which enables the model to learn the potential distribution structure based on limited real channel samples and generate "pseudo-real" channel sequences with high fidelity, thereby effectively expanding the coverage of the training data set and improving the fitting ability of complex multipath, dynamic environment, and high-time-varying channel characteristics.

[0116] In step S2, the goal of the generation model of the embodiment is to generate channel samples with a distribution close to the real channel, that is, to achieve wherein represents the distribution of the real channel data set, represents the distribution of the generated channel data set.

[0117] In the embodiment, the Wasserstein distance is introduced to measure the difference between the real channel distribution and the generated channel distribution as the loss function for training the generation model, which is defined as:

[0118] (16);

[0119] wherein, inf represents the lower bound, represents the mathematical expectation, represents the real data set, represents the generated data set, represents and the set of all joint distributions, denotes the Frobenius norm.

[0120] Based on the Kantorovich-Rubinstein duality, the Wasserstein distance can be rewritten as:

[0121] (17);

[0122] where sup denotes the supremum, Z denotes the set of 1-Lipschitz functions, denotes a real-valued function defined on the channel sample space.

[0123] In some embodiments, the generative model of the present embodiment can employ an adversarial network, including a generator G and a discriminator D; the generator takes a dimensional noise vector z and a scene label c as input, and generates synthetic channel data ; the discriminator takes the scene label c, and the channel impulse response samples of the real channel H or the synthetic channel as input, to distinguish between real and synthetic channel data.

[0124] Further, the objective function of the proposed model is modified as:

[0125] (18);

[0126] where denotes the min-max optimization problem between the generator and the discriminator, that is, the discriminator maximizes the objective function to enhance the ability to distinguish between real and generated data, and the generator minimizes the objective function to make the generated data distribution approximate the real data distribution, respectively denote the distributions of X, z, respectively denote the distributions of X, z, denotes the Euclidean norm , is a random variable subject to a uniform distribution from 0 to 1. In addition, in formula (18), the last term is a gradient penalty term, is a gradient penalty weight coefficient, denotes the gradient operator.

[0127] The generative model of the present embodiment, the generator input includes a latent noise vector and a scene label; the latent noise vector provides randomness for the generation process, and the scene label conditionally controls the generated channel impulse response;

[0128] Further, in the generator, the input scene label and noise vector are processed as follows:

[0129] Step S21, the scene label is embedded by the embedding layer of the generator, the discrete scene label is mapped into a high-dimensional vector to capture the scene semantic information related to the environment, and then the label vector is reshaped into a tensor suitable for time series processing through dimension transformation; and the random vector is taken as a noise vector;

[0130] Step S22, the noise vector is sampled from a standard normal distribution, up-sampled through a fully connected layer to increase the feature dimension, to obtain an expanded noise vector, and then the noise vector is dimensionally transformed to be reshaped into a noise tensor consistent with the shape of the label tensor;

[0131] Step S23, the noise tensor and the label tensor are spliced into a composite feature, which is processed through a multi-layer deconvolution network and a nonlinear activation layer in the time dimension to generate time series data corresponding to each scene;

[0132] The multi-layer deconvolution network and the nonlinear activation layer are a transposed convolution layer and an activation function layer as shown in Figure 3 .

[0133] Step S24, the time series data in each scene is processed in a dimension reduction manner to map the time series data into a channel impulse response sequence of the target scene, i.e., to generate synthesized channel data .

[0134] In this step, a fully connected layer can be used to process the time series data in a dimension reduction manner to map the time series data into a channel impulse response sequence of the target scene.

[0135] Further, in the discriminator, the input scene label c and the channel impulse response sample of the real channel H or the synthesized channel are processed, and the process is as follows:

[0136] Step S201, the scene label is embedded in the embedding layer of the discriminator to be converted into a high-dimensional vector, and is further reshaped into a tensor matching the channel sample, i.e., a label tensor;

[0137] Step S202, the input channel impulse response sample is processed through dimension transformation and a fully connected layer to make it consistent with the dimension of the embedded scene label;

[0138] Step S203, the reshaped channel impulse response sample tensor and the label tensor are spliced in the channel dimension to form a joint feature tensor;

[0139] Step S204, the joint feature tensor is processed through a multi-layer convolutional network to gradually compress the spatial dimension to extract discriminative features. A nonlinear activation function is used after each convolution to enhance the discriminative ability. The final output feature is vectorized and converted into a scalar between 0 and 1 through a fully connected layer, representing the probability that the input time series is real or generated.

[0140] This step is implemented through multiple convolutional layers and activation function layers in the discriminator. Figure 3

[0141] After the output of the discriminator, an adversarial loss function is constructed based on the discriminative outputs of the real channel samples and the generated channel samples, and a gradient penalty term is introduced to constrain the discriminator to satisfy the 1-Lipschitz condition. Subsequently, the network parameters of the discriminator and the generator are alternately optimized through the backpropagation algorithm: the discriminator maximizes the difference between the discriminative outputs of the real samples and the generated samples to enhance its ability to distinguish different channel distributions; the generator minimizes the difference to gradually approach the real channel samples in the discriminator space, thereby promoting the convergence of the synthesized channel data distribution to the real channel data distribution. By alternately training the generator and the discriminator, a generator model capable of generating high-fidelity channel impulse response samples is ultimately obtained.

[0142] In step S3, the enhanced data set is input with the channel impulse response data obtained in the previous time period as input and the channel impulse response at the target time in the future as output. The constructed fusion time-delay AFF and GRU channel prediction model is trained, the channel impulse response data obtained in the previous time period is taken as the input history channel impulse response sequence, steps 2 to 5 are executed, the predicted channel state at the future time is obtained, and compared with the actual channel impulse response at the target time in the future. The loss function value is calculated based on formula (15), and the fusion time-delay AFF and GRU channel prediction model is trained by gradient descent method.

[0143] Further, the hyperparameters set in the training process of the fusion time-delay AFF and GRU channel prediction model include the optimizer, the learning rate, the decay weight, the weight coefficient of the penalty term, the training batch size, etc., which complete the training of the model.

[0144] Based on the real channel and the predicted channel, the key statistical parameters of the channel are calculated, including the power delay spectrum, the root mean square delay spread, and the Rician factor, to verify the accuracy of the fusion time-delay AFF and GRU channel prediction model.

[0145] The root mean square delay spread of the channel is used to measure the dispersion degree of the multipath components in time due to different propagation paths, and its calculation method is:

[0146] ​ (19);

[0147] where m is the delay index, is the average delay, defined as:

[0148] (20);

[0149] denotes the power delay profile, which can be obtained from the channel impulse response samples at Q consecutive measurement instants is estimated as:

[0150] (21);

[0151] K is the Rice factor, which is used to describe the relative strength of the line-of-sight component and the non-line-of-sight component in the wireless channel, and is defined as:

[0152] (22);

[0153] To illustrate the effect of channel prediction in this embodiment, an example simulation verification is carried out, and the specific description is as follows:

[0154] An unmanned aerial vehicle air-to-ground channel measurement platform is built to carry out field measurement to obtain original channel impulse response data;

[0155] In step S1, channel measurement is carried out in a typical campus environment. It covers three types of propagation scenarios to reflect the diversity of channel characteristics. For example, in scenario one, the unmanned aerial vehicle carrying the transmitter hovers above the campus building, and the ground receiver is arranged on the roof of a multi-storey building, about 15 meters high from the ground, and the channel is a typical line-of-sight transmission. Scenario two is the square in front of the library, which is relatively open, and the channel is mainly composed of line-of-sight components with some non-line-of-sight paths. Scenario three is located in the hilly area in the southeast of the campus, which is obviously affected by terrain obstruction and diffraction, and the channel has strong non-line-of-sight components.

[0156] The measurement system consists of an airborne transmitter and a ground receiver. Two software-defined radio (USRP) devices are deployed on the UAV platform and the ground station, respectively, responsible for signal transmission and reception. The transmission and reception processes of both USRPs are uniformly configured and controlled by the GNU Radio platform. At the transmitter, GNU Radio generates and modulates a pseudo-random sequence, selecting an m-sequence with good autocorrelation characteristics as the baseband signal; the signal is then up-converted to radio frequency and transmitted to the ground via the UAV-borne USRP. At the ground receiver, the signal is captured by another USRP and down-converted to baseband, where demodulation and baseband processing are performed in GNU Radio. The system parameters are set as follows: both transmit and receive antenna gain are 30dB, the sampling rate is 25MHz, and the carrier frequency is 2GHz. The UAV flies at an altitude of 20m, with an initial horizontal distance of 50m, and flies along a straight path away from the receiver.

[0157] Received signal It can be represented as a transmitted signal Channel impulse response The convolution of the transmitted signal. The autocorrelation function of the transmitted signal is denoted as... The cross-correlation function between the received and transmitted signals. It can be represented as and The convolution integral form. If we assume the transmitted signal is white noise, and its autocorrelation function is an impulse function, substituting it into the cross-correlation expression yields... Proportional to channel impulse response By normalizing the data, the channel impulse response can be directly estimated from the cross-correlation function.

[0158] Data Augmentation: To address the issues of high cost and strong environmental dependence in channel measurement, this embodiment proposes a data augmentation module. This module effectively expands the limited channel measurement data by constructing a generative adversarial network.

[0159] like Figure 3 As shown, the generator's input is a potential noise vector. and a scene label c. Label c is first mapped to a 256-dimensional vector by an embedding layer and reshaped to a 64x2x2 tensor to provide spatial context information. A noise vector z sampled from a normal distribution is first passed through two fully connected layers to upsample it from 32 dimensions to 256 dimensions, and then reshaped to a 64x2x2 tensor. The label representation and the noise representation are concatenated along the channel dimension to form a 128x2x2 tensor, which is used as the input to subsequent convolutional layers. The concatenated tensor is processed by three transposed convolutional layers, each followed by a LeakyReLU activation function. These transposed convolutional layers upsample the feature maps from 2x2 to 16x16, thereby increasing the spatial resolution of the generated samples. A convolutional layer with a kernel size of 3 and a stride of 1 is used to generate a single-channel output, which is then passed through a LeakyReLU activation function to introduce nonlinearity. The output feature map is flattened and passed through a fully connected layer to reduce its dimensionality to 100 dimensions.

[0160] In the discriminator, the input label c is converted to a 256-dimensional vector by an embedding layer and reshaped to a 16x16 tensor. A synthetic or real sample is reshaped from a 100-dimensional vector to a 16x16 tensor and then passed through a fully connected layer to adjust its dimensionality to match the label embedding representation. The sample tensor and the label embedding are concatenated along the channel dimension to form a 2x16x16 tensor, which is input to a convolutional layer. The tensor is processed by three convolutional layers, each followed by a LeakyReLU activation function, to reduce the spatial dimensions from 16x16 to 2x2, thereby gradually extracting hierarchical features from the input data. The feature map output by the final convolutional layer is flattened and passed through a fully connected layer to output a scalar value representing the probability that the input is a real sample.

[0161] The channel prediction of the fusion time-delay AFF and GRU channel prediction model of the embodiment is as shown in Figure 3 The input is a historical sequence with a dimension of BxLxD, and a normalized target timestamp t with a dimension of Bx1, where B=128 is the batch size, L=10 is the lookback step number, and D=100 is the number of multipaths. First, a delay-domain AFF dictionary matrix of size is constructed for each tap on the normalized delay grid, where is the basis of the delay AFF, and the dictionary is composed of learnable frequency and phase vectors, both of which have a length of . Based on the dictionary, a positive semi-definite weight matrix A with a dimension of is introduced in the feature space, and a learnable span kernel with a size of DxD is constructed on the delay axis to characterize the correlation between multipaths, realize coupling, smoothing, and energy redistribution of different multipaths, and explicitly inject the multipath correlation into the feature representation. Then, the last dimension D of each frame in the historical sequence is left multiplied to obtain a span-mixed sequence with the same size as the original sequence , with dimension B x L x D, which completes the learnable filtering and energy re-allocation parameterized by AFF on the delay axis, making the subsequent time series modeling more robust. The mixed sequence is input into GRU to extract the temporal context representation , with dimension B x H, where H = 256, and the GRU adopts the standard implementation of reset gate, update gate and candidate state.

[0162] In parallel, the time-domain AFF feature is calculated for the target timestamp t, resulting in a vector with dimension , where = 240 is the time AFF base, and both the frequency and the phase are learnable vectors with length ; at the same time, the last frame is projected into the delay AFF dictionary to obtain the delay semantic vector, with dimension . Then the features , the time AFF and the delay semantic are concatenated to form the fusion representation z, with dimension , which is fed into a multi-layer perceptron for nonlinear fusion and prediction. The multi-layer perceptron adopts a three-layer fully connected structure, with channel widths of 512, 512 to 256, and 256 to 128, respectively. Each layer contains a linear mapping, batch normalization and a ReLU activation function. Dropout is applied in the first two layers to suppress overfitting. Finally, a linear mapping is used to output the next frame channel impulse response prediction , with dimension B x D.

[0163] The enhanced channel dataset is used to train and test the model. In the channel generation phase, a total of 3000 sample data are collected, corresponding to three scene labels of 0, 1 and 2, with 1000 samples for each scene. The dataset is divided into training and test sets in the ratio of 8:2. The model is trained for a total of 3000 rounds, and the discriminator and generator are updated at a frequency of 3:1. A gradient penalty mechanism is introduced during the training process, with a penalty coefficient of 10. The learning rate of the generator is set to 3 x 10 -5 , and the learning rate of the discriminator is set to 1 x 10 -4 , both of which are linearly decayed to 0 throughout the training process. The optimizer is Adam. For the channel prediction phase, the sample size of each scene is expanded to 3000, and the training and test sets are also divided in the ratio of 8:2. The time index and feature value of the input data are normalized to the [0, 1] interval using MinMaxScaler. The sample is constructed by sliding window, with a window length of 10 and a sampling interval of 1 second. The historical input sequence, target time index and corresponding next frame target value are generated, and loaded into the data loader with a batch size of 128. The upper limit of the training rounds is 2000, and the optimizer is Adam with an initial learning rate of 1 x 10 -3, the weight decay coefficient is 1 x 10 -4 , the loss function adopts a weighted combination of mean square error and SmoothL1 loss, and the weight coefficient is 0.5.

[0164] In order to further show the effectiveness of the method, Figure 4 The root mean square delay spread of the measured and generated channels in different scenarios is compared. In scenario one, the channel has small delay spread due to strong line-of-sight propagation. In contrast, the delay spread in the other two scenarios increases significantly, especially in scenario three, which exhibits a more complex delay spread structure due to rich multipath propagation. The high consistency between the measured and generated channels using the method of the embodiment shows that the proposed model effectively captures the time dispersion characteristics of the channel under different environmental conditions.

[0165] Figure 5 The Rician K factor of the measured and generated channels in different scenarios is further compared. The Rician K factor of scenario one is about 12dB, indicating that the line-of-sight component dominates. In scenario two, the Rician K factor drops to about 7dB, indicating that the line-of-sight propagation still dominates but the scattering and multipath effects are enhanced. In scenario three, the Rician K factor further decreases to about 3dB, reflecting that the signal experiences significant scattering and multipath propagation. The generated results are highly consistent with the measured data, indicating that the proposed generation model can accurately depict the power ratio relationship between the line-of-sight and multipath components under different channel conditions.

[0166] Table 1 shows the channel prediction performance comparison of the proposed prediction model and existing models, including BPNN (Back Propagation Neural Network) and LSTM (Long Short-Term Memory). The evaluation indicators include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R 2 ). The BPNN adopts a three-layer feedforward structure, with two hidden layers containing 128 and 256 neurons respectively, each followed by a ReLU activation function. Finally, a 256-dimensional latent representation is mapped to a 100-dimensional output to predict the next time channel state. The LSTM adopts a double-layer bidirectional structure, with 256 hidden units in each layer, forming a 512-dimensional hidden representation after splicing. The time series with a window length of 10 are captured for bidirectional temporal dependence, and the average pooling of the final hidden states of all nodes is used to obtain the global features, and then a 100-dimensional vector is output through a fully connected layer. A dropout rate of 0.2 is used in training to prevent overfitting.

[0167] Table 1 Comparison of channel prediction performance in different scenarios

[0168]

[0169] As shown in the results of Table 1, under three different propagation scenarios, the proposed prediction model is superior to the comparative model in terms of mean square error, root mean square error, mean absolute error, and determination coefficient. Compared with the BPNN and LSTM models, the proposed method shows a significant downward trend in error indicators, and the determination coefficient is closer to 1, indicating that the prediction results have higher consistency with the true channel state. The prediction results of the BPNN model deviate greatly in scenarios where the multipath power changes dramatically, because the model does not explicitly model the time correlation of the channel. Although the LSTM model has certain time modeling capability, its ability to describe high-frequency changes and complex multipath structures in the channel is limited, and the prediction accuracy is still restricted. In contrast, the proposed prediction model can more accurately reconstruct the overall distribution and local details of the channel power delay profile, and has higher prediction accuracy and stability under different scenario conditions. In summary, the comparison results in the table verify the effectiveness of the proposed prediction model in complex air-to-ground channel environments, which can achieve high-precision prediction of channel states under multiple scenario conditions, and has good generalization ability and robustness.

[0170] Embodiment 2

[0171] Based on Embodiment 1, an unmanned aerial vehicle-to-ground channel prediction system based on adaptive Fourier features is provided in this embodiment, comprising:

[0172] A measurement data acquisition module configured to acquire an unmanned aerial vehicle-to-ground channel to be predicted and acquire a measured historical channel impulse response sequence;

[0173] An AFF mapping module configured to extract the timestamp and multipath delay corresponding to each historical channel impulse response, respectively perform adaptive Fourier feature mapping, and obtain time AFF mapping and delay AFF mapping;

[0174] A context information extraction module configured to define a delay semantic vector based on the channel impulse response of the previous frame of channel signals and the delay AFF mapping, and extract the context information of the delay of different path signals in multipath propagation;

[0175] A time series encoding module configured to model the correlation between different multipath signals based on the delay AFF mapping, construct a cross-path kernel for the historical channel impulse response sequence mixture, obtain a new channel representation after cross-path, and input it to a GRU time series encoder to extract time context encoding;

[0176] A fusion and regression prediction module configured to fuse the time context encoding, the time AFF mapping, and the context information of the delay to obtain fused spatio-temporal features, and perform regression prediction to obtain the channel state at the future time.

[0177] Further, the time sequence coding module is configured to perform the following steps:

[0178] The time delay AFF mapping features of all the multipaths are stacked into a matrix to obtain a time delay AFF dictionary;

[0179] Based on the time delay AFF dictionary obtained based on the time delay AFF mapping, a learnable inter-path kernel is defined on the time delay axis based on the time delay AFF dictionary to model the correlation between different paths.

[0180] For any historical channel impulse response, a new channel representation, i.e., an inter-path channel representation, is obtained by multiplying the inter-path kernel.

[0181] It should be noted that each module in the embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0182] Embodiment 3

[0183] Based on Embodiment 1, the present embodiment provides a UAV-to-ground channel prediction system based on adaptive Fourier features, comprising a collection device and a processor.

[0184] The collection device is used to collect a historical channel impulse response sequence.

[0185] The processor is configured to perform the steps of the UAV-to-ground channel prediction method based on adaptive Fourier features described in Embodiment 1.

[0186] Specifically, the channel measurement system used by the collection device includes an air signal transmitting unit and a ground receiving unit, and the channel measurement system includes an air signal transmitting unit and a ground receiving unit.

[0187] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0188] The above describes the specific embodiments of the present application in conjunction with the accompanying drawings, but is not intended to limit the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A UAV ground channel prediction method based on adaptive Fourier features, characterized in that, Includes the following steps: Obtain the ground-to-ground channel of the UAV to be predicted, and obtain the measured historical channel impulse response sequence; Extract the timestamp and multipath delay corresponding to each historical channel impulse response, and perform adaptive Fourier feature mapping to obtain the time AFF mapping and the delay AFF mapping; including the following steps: Adaptive Fourier feature mapping is performed on the timestamps corresponding to the channel impulse response to obtain the time-series dynamic change information of the channel, i.e., time AFF mapping; An adaptive Fourier transform is used to map the time delay of the channel impulse response for different paths in multipath propagation, resulting in a time delay AFF mapping. Based on the channel impulse response and delay AFF mapping of the previous frame's channel signal, a delay semantic vector is defined to extract the contextual information of the delay of signals from different paths in multipath propagation. The delay AFF mapping features of all multipath signals are stacked into a matrix to obtain a delay AFF dictionary. Based on the channel impulse response and delay AFF dictionary of the previous frame's channel signal, a delay semantic vector is defined to obtain the contextual information of the delay of signals from different paths in multipath propagation. The formula is as follows: ； in, Represents the delay AFF dictionary, This represents the channel impulse response of the channel signal in the previous frame; Indicates the base of the delay AFF; The correlation between different multipath signals is modeled based on time-delay AFF mapping. A cross-path kernel is constructed and mixed with historical channel impulse response sequences to obtain a new cross-path channel representation. This new representation is then input into a GRU time encoder to extract the time context code. The process of modeling the correlation between different multipath signals based on time-delay AFF mapping and constructing a cross-path kernel to mix with historical channel impulse response sequences to obtain a new cross-path channel representation includes the following steps: Based on the time-delay AFF dictionary obtained from the time-delay AFF mapping, a learnable span kernel is defined on the time-delay axis to model the correlation between different paths, as shown below: ； Here, A is a positive semi-definite matrix. Indicates the transect nucleus. Represents the delay AFF dictionary, Indicates transpose. D It is the number of multipaths; For any historical channel impulse response, a new channel representation is obtained by multiplying it with the crossover kernel, i.e., the channel representation after crossover; By fusing time context encoding, time AFF mapping, and time delay context information, the fused spatiotemporal features are obtained, and regression prediction is performed to obtain the channel state at future moments.

2. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 1, characterized in that: It also includes building a channel prediction model that integrates time-delay AFF and GRU, including: an adaptive Fourier transform module, a GRU timing encoder, and a multilayer perceptron regressor; An adaptive Fourier transform module is configured to extract the timestamp and multipath delay corresponding to each historical channel impulse response, and perform adaptive Fourier feature mapping to obtain a time AFF mapping and a delay AFF mapping. Based on the channel impulse response and delay AFF mapping of the previous frame channel signal, a delay semantic vector is defined to extract the context information of the delay of signals of different paths in multipath propagation. Based on the delay AFF mapping, the correlation between different multipath signals is modeled, and a cross-path kernel is constructed to mix the historical channel impulse response sequences to obtain a new cross-path channel representation. The GRU timing encoder is used to receive the channel representation after cross-path mixing and extract the time context encoding. The multilayer perceptron regressor is used to perform regression prediction on the fused spatiotemporal features to obtain the channel state at future times.

3. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 2, characterized in that, It also includes the process of training the channel prediction model that integrates time-delay AFF and GRU, including the following: Obtain the UAV-to-ground channel to be predicted, and obtain the measured historical channel impulse response sequence; An adversarial learning network is constructed as a generative model to augment the acquired historical channel impulse response sequence data and build a training dataset. Based on the enhanced dataset, the channel impulse response data obtained in the previous time period is used as input, and the channel impulse response at the next target time is used as output. The constructed channel prediction model of fused time-delay AFF and GRU is trained to obtain the trained channel prediction model of fused time-delay AFF and GRU.

4. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 3, characterized in that: An adversarial network consists of a generator G and a discriminator D; the generator uses a multidimensional noise vector. Using scene label c as input, synthetic channel data is generated. The discriminator uses the scene label c and the real channel H or the synthetic channel. Channel impulse response samples are used as input to distinguish between real and fake channel data.

5. A UAV ground-to-ground channel prediction system based on adaptive Fourier features, characterized in that, include: The measurement data acquisition module is configured to acquire the UAV-to-ground channel to be predicted and to acquire the historical channel impulse response sequence of the measurement. The AFF mapping module is configured to extract the timestamp and multipath delay corresponding to each historical channel impulse response, and perform adaptive Fourier feature mapping to obtain the time AFF mapping and the delay AFF mapping; including the following steps: Adaptive Fourier feature mapping is performed on the timestamps corresponding to the channel impulse response to obtain the time-series dynamic change information of the channel, i.e., time AFF mapping; An adaptive Fourier transform is used to map the time delay of the channel impulse response for different paths in multipath propagation, resulting in a time delay AFF mapping. The context information extraction module is configured to extract the context information of the time delay of signals from different paths in multipath propagation by defining a time delay semantic vector based on the channel impulse response and time delay AFF mapping of the channel signal from the previous frame; stacking the time delay AFF mapping features of all multipaths into a matrix to obtain a time delay AFF dictionary; and defining a time delay semantic vector based on the channel impulse response and time delay AFF dictionary of the channel signal from the previous frame to obtain the context information of the time delay of signals from different paths in multipath propagation, as shown in the formula: ； in, Represents the delay AFF dictionary, This represents the channel impulse response of the channel signal in the previous frame; Indicates the base of the delay AFF; The timing coding module is configured to model the correlation between different multipath signals based on the time-delay AFF mapping, construct a cross-path kernel, mix historical channel impulse response sequences to obtain a new cross-path channel representation, and input it into the GRU timing encoder to extract the time context code. The process of modeling the correlation between different multipath signals based on the time-delay AFF mapping and constructing a cross-path kernel to mix historical channel impulse response sequences to obtain a new cross-path channel representation includes the following steps: Based on the time-delay AFF dictionary obtained from the time-delay AFF mapping, a learnable span kernel is defined on the time-delay axis to model the correlation between different paths, as shown below: ； Here, A is a positive semi-definite matrix. Indicates the transect nucleus. Represents the delay AFF dictionary, Indicates transpose. D It is the number of multipaths; For any historical channel impulse response, a new channel representation is obtained by multiplying it with the crossover kernel, i.e., the channel representation after crossover; The fusion and regression prediction module is configured to fuse time context encoding, time AFF mapping, and time delay context information to obtain fused spatiotemporal features, and then perform regression prediction to obtain the channel state at future times.

6. A UAV ground-to-ground channel prediction system based on adaptive Fourier features, characterized in that, include: Data acquisition device and processor; Acquisition device, used to acquire historical channel impulse response sequences; The processor is configured to perform the steps of the UAV ground channel prediction method based on adaptive Fourier features as described in any one of claims 1-4.

Images