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

Abstract

The invention relates to the technical field of channel modeling and prediction, and provides an unmanned aerial vehicle to ground channel prediction method and system based on adaptive Fourier features, and the method comprises the steps: obtaining a historical channel impact response sequence, extracting a timestamp and a multipath time delay, carrying out the adaptive Fourier feature mapping, and obtaining a channel impact response sequence; identifying context information of time delays of signals of different paths in multipath propagation; modeling correlation among different multipath signals based on time delay AFF mapping, constructing span core historical channel impact response sequence mixing, and extracting time context codes; and fusing the context information of the time context coding, the time AFF mapping and the time delay, and performing regression prediction to obtain a channel state at a future moment. According to the method, the adaptive Fourier feature mapping, the time delay semantic vector and the span kernel mechanism are introduced, and the GRU time sequence coding and the multi-feature fusion strategy are fused, so that accurate modeling and efficient prediction of the unmanned aerial vehicle to-ground channel rapid time-varying feature are realized.

Description

Technical Field

[0001] This invention relates to the field of channel modeling and prediction technology, specifically to a method and system for predicting UAV ground channels based on adaptive Fourier features. Background Technology

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

[0003] With the large-scale deployment of drone technology in civilian, industrial, and public safety fields, air-to-ground wireless communication has become a core infrastructure for ensuring the safe flight and mission execution of drones. When drones perform tasks such as aerial photography and inspection, logistics transportation, disaster assessment, and communication relay, they must rely on stable and reliable wireless links to issue remote control commands, monitor flight status, and transmit service data. However, compared with traditional terrestrial mobile communication, drones have characteristics such as high flight altitude, high speed, and frequent environmental changes. This causes their air-to-ground signals to be affected by a combination of factors during propagation, including path loss, multipath reflection, ground obstruction, and Doppler shift, resulting in highly dynamic, highly random, and highly time-varying channel characteristics.

[0004] In complex airspace environments, the communication reliability of UAVs depends not only on the current channel quality but also on the communication system's ability to predict future channel changes. Accurate channel modeling and real-time channel estimation provide the communication system with a characterization and prediction of the wireless propagation environment, enabling the system to anticipate link quality changes and adopt more robust communication strategies. For example, by predicting through channel modeling that a UAV might enter a weak coverage area, the system can adjust its transmit power, switch link modes, or adjust modulation and coding schemes in advance. Furthermore, channel modeling plays a crucial role in communication resource allocation. Due to the increasing number of UAVs and the diversification of service demands, ground control terminals or cellular base stations need to allocate spectrum, time slots, or power resources to different UAVs based on channel prediction results to improve overall network throughput and link stability. Therefore, channel models are not only tools for describing physical propagation characteristics but also important bases for intelligent scheduling and resource optimization in communication systems. They realize the core logical chain of "physical channel modeling → wireless environment prediction → resource allocation strategy optimization," directly affecting the reliability, real-time performance, and spectrum utilization efficiency of UAV communication networks.

[0005] Although the rapid development of wireless communication and artificial intelligence technologies in recent years has provided new methods for UAV channel modeling, existing channel modeling and prediction research still has many shortcomings. Regarding the modeling objects, most current channel research focuses on terrestrial cellular communication scenarios, while systematic research on UAV air-to-ground channels is relatively limited. At the channel characteristics level, the received signal power during UAV flight is affected by multipath propagation and fading in the time delay domain and by the Doppler effect in the time domain, exhibiting strong randomness and rapid changes. The channel impulse response contains a large number of high-frequency components on both the time axis and the multipath delay axis. Traditional neural network-based prediction methods are limited by the inherent spectral bias problem, often learning low-frequency variation characteristics more easily and failing to effectively characterize the high-frequency time-varying behavior in UAV channels, resulting in insufficient modeling and prediction accuracy for rapidly time-varying channels. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method and system for predicting UAV ground-to-ground channels based on adaptive Fourier features. By introducing adaptive Fourier feature mapping, time delay semantic vectors, and a cross-path kernel mechanism, and integrating GRU time-series coding and multi-feature fusion strategies, it achieves accurate modeling and efficient prediction of the rapidly changing time-varying characteristics of UAV ground-to-ground channels.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for predicting UAV ground channels based on adaptive Fourier features, comprising the following steps: Obtain the UAV-to-ground channel 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 respectively; Based on the channel impulse response of the previous frame channel signal and the time delay AFF mapping, a time delay semantic vector is defined to extract the time delay context information of signals from different paths in multipath propagation. Based on the time-delay AFF mapping model, the correlation between different multipath signals is modeled, the cross-path kernel is constructed and the historical channel impulse response sequence is mixed to obtain a new cross-path channel representation, which is then input into the GRU time encoder to extract the time context code. 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.

[0008] A second aspect of the present invention provides a UAV ground channel prediction system based on adaptive Fourier features, comprising: 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. The context information extraction module is configured to extract the time delay context information of signals from different paths in multipath propagation based on the channel impulse response and time delay AFF mapping definition of the channel signal of the previous frame. The timing coding module is configured to model the correlation between different multipath signals based on the time-delay AFF mapping, construct the cross-path kernel, mix the historical channel impulse response sequence, obtain a new cross-path channel representation, input it to the GRU timing encoder, and extract the time context code. 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.

[0009] A third aspect of the present invention provides a UAV ground channel prediction system based on adaptive Fourier features, comprising: a data acquisition device and a 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 described above.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: The method of this invention employs AFF mapping to effectively address the high-frequency oscillation characteristics in the time and delay domains of UAV channels, enhancing the model's ability to perceive rapidly time-varying channels and solving the problem that traditional models struggle to capture high-frequency changes. By constructing delay semantic vectors and cross-path kernel functions, the method enhances the model's ability to model multipath structural changes, enabling a more comprehensive capture of semantic dependencies and dynamic associations between paths, overcoming the limitations of existing models in dynamic multipath modeling. The introduction of a GRU encoder to capture temporal context information further improves the temporal consistency and accuracy of predictions, effectively enhancing the coherence of predictions at multiple time points. Multi-feature fusion improves the spatiotemporal feature representation capabilities, giving the model stronger generalization and robustness in complex scenarios, thus making it applicable to various environments and task scenarios and alleviating the limitations of existing methods that rely on specific scenarios.

[0011] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 This is a flowchart of the UAV ground channel prediction method based on adaptive Fourier features according to Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating the UAV ground channel prediction method based on adaptive Fourier features according to Embodiment 1 of the present invention. Figure 3 This is a structural diagram of the UAV channel data enhancement module based on generative adversarial networks in Embodiment 1 of the present invention; Figure 4 This is a comparison chart of the generated root mean square delay spread of the channel and the actual root mean square delay spread of the channel under different scenarios in the simulation example of Embodiment 1 of the present invention; Figure 5 This is a comparison chart of the channel Rice factor generated in different scenarios and the actual channel Rice factor in the simulation example of Embodiment 1 of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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 invention pertains.

[0016] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0017] Explanation of technical terms: Channel: refers to the comprehensive manifestation of all physical processes and propagation environment experienced by a wireless signal during propagation. It is used to characterize effects such as amplitude attenuation, phase rotation, frequency shift, time delay spread, and random fading during the transmission of electromagnetic waves from the transmitter to the receiver. Channel models describe these propagation characteristics and are the foundation for communication system performance analysis, link prediction, and parameter optimization.

[0018] Spectral bias: When learning function mappings, neural networks tend to preferentially fit low-frequency components while being relatively inadequate in fitting high-frequency components. In channel prediction tasks, spectral bias makes the model more likely to learn smooth, slow-changing trends in channel variations, while making it difficult to accurately characterize the rapid fluctuations and high-frequency changes caused by high-speed motion, Doppler effects, and complex multipath propagation.

[0019] The time domain refers to the dimension that uses time as the independent variable to describe the channel characteristics, and is used to characterize the dynamic characteristics of the wireless channel as it evolves over time. In UAV communication scenarios, the time domain mainly reflects the rapid changes in channel behavior over time caused by factors such as UAV movement, Doppler shift, and environmental changes.

[0020] The time delay domain refers to the dimension that describes channel characteristics using signal propagation delay as the independent variable. It is used to characterize the multipath propagation structure and the relative delay characteristics of each propagation path in a wireless channel. In UAV air-to-ground channels, the time delay domain reflects the arrival time differences and energy distribution of different multipath components, and is an important foundation for characterizing multipath fading and power delay spectrum.

[0021] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 5 As shown, a UAV ground channel prediction method based on adaptive Fourier features includes the following steps: Step 1: Obtain the UAV-to-ground channel to be predicted and acquire the measured historical channel impulse response sequence; Step 2: Extract the timestamp corresponding to each historical channel impulse response. and multipath delay Adaptive Fourier feature mapping (AFF mapping) is performed on each feature to obtain the temporal AFF mapping. Mapping with delay AFF ; Step 3: Based on the channel impulse response and delay AFF mapping of the previous frame's channel signal, define the delay semantic vector and extract the contextual information of the delay of signals from different paths in multipath propagation. ; Step 4: Based on the time-delay AFF mapping, model the correlation between different multipath signals, construct a cross-path kernel, and mix the historical channel impulse response sequences to obtain a new cross-path channel representation. The input is fed into the GRU (Gated Recurrent Unit) timing encoder to extract the time context code. ; Step 5: Encode the time context Time AFF mapping and contextual information about latency The spatiotemporal features are fused to obtain the fused features, and regression prediction is performed to obtain the channel state at future moments. In this embodiment, the time-varying characteristics of the UAV-to-ground wireless channel are modeled based on Adaptive Fourier Feature (AFF). First, the timestamp and delay parameters are extracted by measuring historical channel impulse responses. For the high-frequency oscillation characteristics in the time and delay domains, adaptive Fourier feature mapping is performed on the timestamp t and delay τ respectively, thereby obtaining more representative frequency domain features. Next, using the channel state and delay mapping from the previous frame, a delay semantic vector is constructed to explicitly capture the temporal structure relationships of different paths in the multipath signal. Simultaneously, 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 to a GRU time-series encoder to further extract dynamic context features from the time series. Finally, the fused features are jointly input into a regression network to predict the channel state at future moments, achieving accurate modeling of the high-speed dynamic channel.

[0022] This embodiment employs AFF mapping to effectively address the high-frequency oscillation characteristics in the time and delay domains of UAV channels, enhancing the model's ability to perceive rapidly time-varying channels and solving the problem of traditional models' difficulty in capturing high-frequency changes. By constructing delay semantic vectors and cross-path kernel functions, the model's ability to model multipath structural changes is enhanced, enabling a more comprehensive capture of semantic dependencies and dynamic associations between paths, overcoming the limitations of existing models in dynamic multipath modeling. The introduction of a GRU encoder to capture temporal context information further improves the temporal consistency and accuracy of predictions, effectively enhancing the coherence of predictions at multiple time points. Multi-feature fusion improves the spatiotemporal feature representation capabilities, giving the model stronger generalization and robustness in complex scenarios, thus making it applicable to various environments and task scenarios and alleviating the limitations of existing methods that rely on specific scenarios.

[0023] To address the rapid changes in channel characteristics in the time and delay domains, this embodiment introduces adaptive Fourier feature mapping in both time and delay dimensions. This maps time and delay parameters to a high-dimensional, learnable frequency domain feature space, effectively alleviating the problem of insufficient modeling ability of traditional neural networks for high-frequency components. It also enhances the model's ability to represent rapid changes caused by the Doppler effect and multipath fading, thereby achieving accurate modeling and prediction of the rapid time-varying characteristics of UAV air-to-ground channels.

[0024] The above steps are achieved by constructing a channel prediction model that integrates time-delay AFF and GRU. The model also includes: an adaptive Fourier transform module, a GRU (Gated Recurrent Unit) timing encoder, and a multilayer perceptron regressor. The adaptive Fourier transform module is configured to extract the timestamp t and multipath delay τ corresponding to each historical shock response and perform adaptive Fourier feature mapping (AFF mapping) to obtain the time AFF mapping. Mapping with delay AFF Based on the channel signal of the previous frame, a time delay semantic vector is defined to extract the contextual information of the time delay of signals from different paths in multipath propagation. Based on the time-delay AFF mapping model, the correlation between different multipath signals is modeled, and a cross-path kernel is constructed to mix historical channel impulse response sequences to obtain a new cross-path channel representation. ; A GRU (Gated Recurrent Unit) timing encoder is used to receive the channel representation after cross-path mixing and extract the time context encoding. A multilayer perceptron regressor is used to perform regression prediction on the fused spatiotemporal features to obtain the channel state at future moments, such as the channel impulse response. Furthermore, it also includes adversarial networks for data augmentation of the acquired training data; In step 1, the acquisition of the historical channel impulse response sequence in this embodiment utilizes a channel measurement system comprising an airborne signal transmitting unit and a ground receiving unit. In practical applications, for UAV communication scenarios, the UAV's signal transmitting end is the airborne signal transmitting unit. The transmitting and receiving units employ software-defined radio devices, which are uniformly configured and coordinated through the GNU Radio platform. The system transmits a pseudo-random sequence with good autocorrelation characteristics, up-converted to the carrier frequency, and then transmitted by the UAV. The ground receiving unit down-converts the signal and performs baseband processing, then extracts the channel response through correlation calculations. Multi-band UAV channel measurements are conducted by setting the transmit and receive antenna gain, system sampling rate, UAV flight altitude, initial horizontal distance, and flight trajectory.

[0025] In step 1, the received and transmitted signals are obtained and estimated to obtain the channel impulse response. The process is as follows: Step 11: Receive the signal By transmitting signals Channel impulse response The formula is as follows: (1); in, Indicates the time delay, where t represents time; Step 12: Transmit the signal As white noise, the autocorrelation of the transmitted signal is calculated using the following function: (2); in, For white noise power density, For the Dirac function, To transmit signals The autocorrelation function; Step 13: Calculate the received signal based on the autocorrelation of the transmitted signal. With transmitted signal The cross-correlation is expressed as follows: (3); in, This represents the convolution operation. , These represent the integration variable and the time delay, respectively.

[0026] Substituting equation (2) into equation (3), based on the assumption that the transmitted signal is white noise, its autocorrelation function is the impulse function. After substituting into the cross-correlation expression, and utilizing the sampling characteristics of the Dirac function, we can obtain that the cross-correlation function between the received signal and the transmitted signal is proportional to the channel impulse response, thereby achieving the estimation of the channel impulse response. We can obtain: (4); Step 14: Based on cross-correlation, the channel impulse response is obtained, which can be estimated as follows: (5); The method in this embodiment utilizes the good correlation characteristics of pseudo-random sequences to achieve effective estimation of channel impulse response in complex air-to-ground environments.

[0027] In step 2, AFFs in the time and delay domains are introduced to capture timing variations and frequency-selective fading in the wireless channel.

[0028] In step 2, the timestamp corresponding to each historical channel impulse response is extracted. and multipath delay Adaptive Fourier feature mapping (AFF mapping) is performed on each feature to obtain the temporal AFF mapping. Mapping with delay AFF The method includes the following steps: Step 21, Time AFF Mapping: 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. ; For normalized time Time AFF mapping Defined as: (6); in, As the base of time AFF, These represent the learnable frequency and phase of the time-dependent AFF, respectively. 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; For each normalized delay The delay AFF mapping is defined as follows: (7); in, The base of the delay AFF is represented. These represent the learnable frequency and phase of the delay AFF, respectively; 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. Step 3, the method for extracting the contextual information of the time delay of signals from different paths in multipath propagation, includes: Step 31: Stack the delay AFF mapping features of all multipaths into a matrix to obtain the delay AFF dictionary, represented as: (8); 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; 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. 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. ; (9); in, This represents the transpose of the delay AFF dictionary. This represents the channel impulse response of the channel signal in the previous frame; The base of the delay AFF; In this step, prior information about the delay is provided by projecting the channel impulse response of the previous frame, which helps the model to perform timing modeling.

[0029] In step 4, the correlation between different multipath signals is modeled based on the time-delay AFF mapping, and a cross-path kernel is constructed to mix the historical channel impulse response sequences to obtain a new cross-path channel representation. It includes the following steps: Step 41: Based on the time delay AFF mapping obtained in Step 31, a learnable span kernel is defined on the time delay axis to model the correlation between different paths, as shown below: (10); Here, A is a positive semi-definite matrix representing the structure of the span kernel, which can be chosen in diagonal form to reduce computational complexity. Indicates the transect nucleus. A dictionary representing delay AFFs, Indicates transpose. D It is the number of multipaths; through the transpath nucleus The model can learn the correlation between different multipaths, thereby enhancing its robustness to multipath interference.

[0030] Each sampling point of the channel impulse response corresponds to the time delay of a different propagation path in the wireless channel; arranging these time delays in chronological order forms the time delay axis.

[0031] Step 42: Response to any historical channel impulse By multiplying with the crossover kernel, a new channel representation is obtained, which is the sequence of historical channel response representations after the crossover: (11); Where L is the length of the historical window, representing the size of the time series window used for prediction.

[0032] The above implementation method in this embodiment, by introducing a time delay AFF mapping and a learnable crosspath kernel on the time delay axis, can explicitly model the structured correlation between multipaths, thereby significantly improving the accuracy and robustness of channel impulse response prediction. First, the time delay AFF maps the scalar time delay to a continuous high-dimensional learnable feature space, making the multipath time delay distribution form a captureable geometric structure within the feature domain. Then, the crosspath kernel constructed through the time delay AFF dictionary matrix and the kernel matrix can effectively describe the degree of correlation between different multipath paths, achieving a weighted mixing of multipath gains. This process is equivalent to introducing a multipath structure prior to the historical channel sequence, enabling the channel representation after crosspath kernel processing to simultaneously possess physical meaning and stronger separability.

[0033] In step 4, further, the channel representation after the crossing is... The input is fed into the GRU (Gated Recurrent Unit) timing encoder to extract the time context code. The formula is: (12); in, This represents a gated loop unit used to extract time context codes. Represents the set of parameters of the GRU; In step 5, the temporal context encoding, temporal AFF mapping, and temporal delay semantic vector are fused to obtain the fused spatiotemporal features, as shown in the formula: (13); The fused spatiotemporal features are input into the multilayer perceptron regressor to obtain the predicted value of the channel impulse response at the next time step: (14); Furthermore, the channel prediction model that integrates time-delay AFF and GRU is trained with the objective of minimizing the weighted L2 loss and SmoothL1 loss. The overall loss function is expressed as follows: (15); in, It is a weighting factor used to balance the two loss functions.

[0034] Furthermore, it also includes the process of training the channel prediction model that integrates time-delay AFF and GRU, including the following: Step S1: Obtain the ground channel of the UAV to be predicted and obtain the measured historical channel impulse response sequence; Step S2: Construct an adversarial learning network as a generative model to augment the acquired historical channel impulse response sequence data and construct a training dataset; Step S3: Based on the enhanced dataset, take the channel impulse response data obtained in the previous time period as input and the channel impulse response at the next target time as output, and train the constructed channel prediction model of fusion time-delay AFF and GRU to obtain the trained channel prediction model of fusion time-delay AFF and GRU. The training method in this embodiment uses an adversarial learning network to augment historical channel impulse response sequences, effectively addressing the common problem of insufficient training data in deep learning channel prediction. Traditional methods typically rely on data generated by simulators or ray tracing for training. While these methods offer controllability and repeatability, they still exhibit significant distributional differences compared to real air-to-ground scenarios, resulting in limited model generalization ability. Furthermore, real channel measurements are constrained by factors such as expensive equipment, high measurement costs, and complex deployment, making it difficult to obtain a sufficient and diverse range of training samples. This embodiment introduces an adversarial learning network, enabling the model to learn the potential distribution structure based on limited measured channel samples and generate high-fidelity "pseudo-measured" channel sequences. This effectively expands the coverage of the training dataset and improves the fitting ability to complex multipath, dynamic environments, and highly time-varying channel characteristics.

[0035] In step S2, the goal of the generative model in this embodiment is to generate channel samples whose distribution closely approximates the real channel, i.e., to achieve... ,in This represents the distribution of the real channel dataset. This indicates the distribution of the generated channel dataset.

[0036] In this embodiment, the Wasserstein distance is introduced to measure the difference between the real channel distribution and the generated channel distribution, and is used as the loss function for training the generative model. Its definition is: (16); Where inf represents the infimum. Represents the mathematical expectation. Represents a real dataset, This indicates the generation of a dataset. express and The set of all joint distributions, This represents the Frobenius norm.

[0037] Based on the Kantorovich-Rubinstein duality, the Wasserstein distance can be rewritten as: (17); Where supremum represents the supremum, and Z represents the set of 1-Lipschitz functions. This represents a real-valued function defined on the channel sample space.

[0038] In some embodiments, the generative model of this embodiment may employ an adversarial network, including a generator G and a discriminator D; the generator uses a... Using a noise vector z and a 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.

[0039] Furthermore, the objective function of the proposed model is modified as follows: (18); in, This represents a minimax optimization problem between a generator and a discriminator. The discriminator enhances its ability to distinguish between real and generated data by maximizing the objective function, while the generator minimizes the objective function to make the distribution of generated data approximate the distribution of real data. Representing X, z, respectively The distribution, Describing the Euclidean norm , Let be a random variable that follows a uniform distribution from 0 to 1. Furthermore, in formula (18), the last term is the gradient penalty term. The gradient penalty weight coefficient, This represents the gradient operator.

[0040] In this embodiment, the generator input includes a latent noise vector and a scene label; the latent noise vector provides randomness to the generation process, while the scene label provides conditional control over the generated channel impulse response. Furthermore, in the generator, the input scene labels and noise vectors are processed as follows: Step S21: Embed the scene labels through the generator's embedding layer to map the discrete scene labels into high-dimensional vectors to capture scene semantic information related to the environment; then, reshape the label vectors into a tensor adapted to temporal processing through dimensional transformation; and use random vectors as noise vectors. Step S22: Sample the noise vector from the standard normal distribution, upsample it through a fully connected layer to increase its feature dimension, and obtain the expanded noise vector. Then, perform a dimension transformation on the noise vector to reshape it into a noise tensor with the same shape as the label tensor. Step S23: The noise tensor and the label tensor are concatenated through channels to form a composite feature, which is then processed through multiple deconvolutional networks and nonlinear activation layers in the time dimension to generate time-series data corresponding to each scene. Multilayer deconvolutional networks and nonlinear activation layers, such as Figure 3 The transposed convolutional layer and activation function layer are shown; Step S24: Perform dimensionality reduction processing on the time-series data for each scenario, mapping the time-series data to the channel impulse response sequence of the target scenario, thus generating synthetic channel data. ; In this step, a fully connected layer can be used to reduce the dimensionality of the time series data, mapping the time series data to a channel impulse response sequence of the target scenario.

[0041] Furthermore, in the discriminator, for the input scene label c, and the real channel H or the synthetic channel... The channel impulse response samples are processed as follows: Step S201: Embed the scene labels in the embedding layer of the discriminator, convert them into high-dimensional vectors, and further reshape them into tensors that match the channel samples, i.e., label tensors; Step S202: The input channel impulse response sample is processed by dimensional transformation and a fully connected layer to make its dimension consistent with that of the scene label after embedding. Step S203: Concatenate the reshaped channel impulse response sample tensor and the label tensor along the channel dimension to form a joint feature tensor; Step S204: The joint feature tensor is processed through a multi-layer convolutional network to progressively compress the spatial dimension in order to extract discriminative features. After each convolutional layer, non-linear activation is used to enhance the discriminative ability. The final output features are vectorized and converted into a scalar of 0 to 1 through a fully connected layer, representing the probability that the input time series is real or generated.

[0042] This step is done Figure 3 Implemented using multiple convolutional layers and activation function layers; After the discriminator outputs its results, adversarial loss functions are constructed based on the discriminant outputs of both real and generated channel samples. A gradient penalty term is introduced to constrain the discriminator to satisfy the 1-Lipschitz condition. Subsequently, the network parameters of the discriminator and generator are alternately optimized using the backpropagation algorithm: the discriminator enhances its ability to distinguish different channel distributions by maximizing the difference between the discriminant outputs of real and generated samples; the generator minimizes the difference, causing the generated channel samples to gradually approach the real channel samples in the discriminator space, thereby promoting the convergence of the synthetic channel data distribution to the real channel data distribution. By alternately training the generator and discriminator, a generative model capable of generating high-fidelity channel impulse response samples is finally obtained.

[0043] In step S3, the enhanced dataset is used as input to the channel impulse response data obtained in the previous time period and as output to the channel impulse response at the next target time. The channel prediction model of the fusion time-delay AFF and GRU is trained. The historical channel impulse response sequence is used as input to the channel impulse response data obtained in the previous time period. Steps 2 to 5 are executed to obtain the predicted channel state at the next time. The predicted channel state is compared with the actual channel impulse response at the next target time. The loss function value is calculated based on formula (15). The channel prediction model of the fusion time-delay AFF and GRU is trained by gradient descent. Furthermore, the hyperparameters set during the training process of the channel prediction model that integrates time-delay AFF and GRU, including optimizer, learning rate, attenuation weights, weight coefficients of penalty terms, training batch size, etc., are used to complete the training of the model.

[0044] Based on real and predicted channels, key statistical parameters of the channel are calculated, including power delay spectrum, root mean square delay spread, and Rice factor, to verify the accuracy of the channel prediction model that fuses time-delay AFF and GRU. The root mean square delay spread of a channel measures the temporal dispersion of multipath components due to different propagation paths. It is calculated as follows: (19); Where m is the delay index. The average delay is defined as: (20); The power delay spectrum can be represented by channel impulse response samples obtained from Q consecutive measurement times. The estimate is as follows: (twenty one); The Rice factor, used to describe the relative intensity of line-of-sight and non-line-of-sight components in a wireless channel, is defined as follows: (twenty two); To illustrate the effectiveness of channel prediction in this embodiment, a simulation was conducted, which is described in detail below: A UAV air-to-ground channel measurement platform was built to conduct field measurements in order to obtain raw channel impulse response data. In step S1, channel measurements are conducted in a typical campus environment. This covers three propagation scenarios to reflect the diversity of channel characteristics. In scenario one, a drone carrying a transmitter hovers over a building on campus, while a ground receiver is deployed on the roof of a multi-story building, with the antenna approximately 15 meters above the ground; the channel exhibits typical line-of-sight transmission. Scenario two is the plaza in front of the library, a relatively open environment; the channel is predominantly line-of-sight, accompanied by several non-line-of-sight paths. Scenario three is located in the hilly southeastern part of the campus, significantly affected by terrain obstruction and diffraction; the channel exhibits strong non-line-of-sight components.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] like Figure 3 As shown, the generator's input is a potential noise vector. A scene label *c* is generated. Label *c* is first mapped to a 256-dimensional vector through an embedding layer and then reshaped into a 64×2×2 tensor to provide spatial context. A noise vector *z* sampled from a normal distribution is first passed through two fully connected layers, upsampling it from 32 dimensions to 256 dimensions, and then reshaped into a 64×2×2 tensor. The label representation and the noise representation are concatenated along the channel dimension to form a 128×2×2 tensor, which serves as the input to subsequent convolutional layers. This concatenated tensor is processed through three transposed convolutional layers, each followed by a LeakyReLU activation function. These transposed convolutional layers upsample the feature map from 2×2 to 16×16, thereby improving the spatial resolution of the generated samples. Finally, a single-channel output is generated using a convolutional layer with a kernel size of 3 and a stride of 1, and then non-linearity is introduced by the LeakyReLU activation function. The output feature map is flattened and then reduced to 100 dimensions through a fully connected layer.

[0049] In the discriminator, the input label *c* is converted into a 256-dimensional vector through an embedding layer and then reshaped into a 16×16 tensor. Synthetic or real samples are reshaped from 100-dimensional vectors into 16×16 tensors, and then their dimensions are adjusted by a fully connected layer to match the label embedding representation. The sample tensor and label embedding are concatenated along the channel dimension to form a 2×16×16 tensor, which is then input into a convolutional layer. This tensor is processed through three convolutional layers, each followed by a LeakyReLU activation function, reducing the spatial size from 16×16 to 2×2, thereby progressively extracting hierarchical features from the input data. Finally, the feature map output from the convolutional layers is flattened and then passed through a fully connected layer to output a scalar value representing the probability that the input is a real sample.

[0050] Channel prediction using the channel prediction model that combines time-delay AFF and GRU in this embodiment, such as Figure 3 As shown, the input is a historical sequence. The dimensions are B×L×D, and the normalized target timestamp t has a dimension of B×1, where B=128 is the batch size, L=10 is the number of lookback steps, and D=100 is the number of multipaths. First, a time-delay grid is constructed for each tap. The time-delay domain AFF dictionary matrix, where , where is the cardinality of the delay AFF, and the dictionary consists of learnable frequency and phase vectors, each with a length of . Based on this dictionary, a dimension of is introduced into the feature space. Determine the positive semidefinite weight matrix A, and construct a learnable span kernel on the time delay axis. The dimension D×D is used to characterize the correlation between multipaths, enabling coupling, smoothing, and energy redistribution of different multipaths, and explicitly injecting multipath correlation into the feature representation. Subsequently, the last dimension D of each frame in the historical sequence is left-multiplied to obtain a cross-path mixing sequence of the same size as the original sequence. The dimension is B×L×D, which represents the learnable filtering and energy redistribution parametrically configured by AFF on the time delay axis, making subsequent time series modeling more robust. The mixed sequence is then input into a GRU to extract the temporal context representation. The dimension is B×H, where H=256. GRU uses the standard implementation of reset gate, update gate and candidate state.

[0051] In parallel, the time-domain AFF feature is calculated for the target timestamp t, and the dimension is obtained. The vector, where =240 is the time AFF base, and both frequency and phase are length. Learnable vectors; and the most recent frame Projecting the vector onto the delay AFF dictionary yields a delay semantic vector, with dimensions [dimension 1]. Subsequently, in terms of features The fused representation z is formed by concatenating time AFF and time delay semantics, with dimension being The data is then fed into a multilayer perceptron for nonlinear fusion and prediction. The multilayer perceptron employs a three-layer fully connected structure, with channel widths of [insert widths here]. Layers 512, 512 to 256, and 256 to 128 each contain linear mapping, batch normalization, and ReLU activation functions. Dropout is applied to the first two layers to suppress overfitting. Finally, a linear mapping layer is used to output the predicted channel impulse response for the next frame. The dimension is B×D.

[0052] The model was trained and tested using an enhanced channel dataset. During the channel generation phase, 3000 sample data points were collected, corresponding to three scenarios labeled 0, 1, and 2, with 1000 samples for each scenario. The dataset was divided into training and testing sets in an 8:2 ratio. The model was trained for 3000 epochs, with the discriminator and generator updating at a 3:1 ratio. A gradient penalty mechanism with a penalty coefficient of 10 was introduced during training. The generator learning rate was set to 3 × 10⁻⁶. -5 The discriminator learning rate is set to 1×10. -4 Both decay linearly to 0 throughout the training process. Adam was used as the optimizer for all scenarios. For the channel prediction stage, the sample size for each scenario was expanded to 3000 samples, and the training and test sets were divided in an 8:2 ratio. The time index and feature values ​​of the input data were uniformly normalized to the [0,1] interval using MinMaxScaler. Samples were constructed using a sliding window method with a window length of 10 and a sampling interval of 1 second, generating historical input sequences, target time indices, and corresponding target values ​​for the next frame, which were then loaded into the data loader in batches of 128. The maximum number of training epochs was 2000, Adam was used as the optimizer, and the initial learning rate was 1×10⁻⁶. -3The weight decay coefficient is 1×10 -4 The loss function uses a weighted combination of mean squared error and SmoothL1 loss, with a weighting coefficient of 0.5.

[0053] To further demonstrate the effectiveness of the proposed method, Figure 4 The root mean square (RMS) delay spread of the measured and generated channels were compared under different scenarios. In scenario one, the channel exhibits a relatively small delay spread due to strong line-of-sight propagation. In contrast, the delay spread increases significantly in the other two scenarios, especially in scenario three, which displays a more complex delay spread structure due to abundant multipath propagation. The high agreement between the measured and generated channels using the method in this embodiment demonstrates that the proposed model effectively captures the temporal dispersion characteristics of the channel under different environmental conditions.

[0054] Figure 5 The Rice K-factors of the measured and generated channels were further compared under different scenarios. In Scenario 1, the Rice K-factor was approximately 12 dB, indicating that the line-of-sight component was dominant. In Scenario 2, the Rice K-factor decreased to approximately 7 dB, indicating that line-of-sight propagation still dominated, but scattering and multipath effects were enhanced. In Scenario 3, the Rice K-factor further decreased to approximately 3 dB, reflecting significant scattering and multipath propagation of the signal. The generated results were highly consistent with the measured data, demonstrating that the proposed generation model can accurately characterize the power ratio relationship between line-of-sight and multipath components under different channel conditions.

[0055] Table 1 shows a comparison of the channel prediction performance of the proposed prediction model with existing models, including BPNN (Back Propagation Neural Network) and LSTM (Long Short-Term Memory). Evaluation metrics include mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The BPNN employs a three-layer feedforward structure, with two hidden layers containing 128 and 256 neurons respectively. Each layer is followed by a ReLU activation function, ultimately mapping the 256-dimensional latent representation to a 100-dimensional output to predict the channel state at the next time step. The LSTM uses a two-layer bidirectional structure, with 256 hidden units per layer, concatenating them to form a 512-dimensional hidden representation. Bidirectional temporal dependency capture is performed on time series with a window length of 10. Average pooling is applied to the final hidden states of all nodes to obtain global features, which are then passed through a fully connected layer to output a 100-dimensional vector. A dropout rate of 0.2 is used during training to prevent overfitting.

[0056] Table 1. Comparison of channel prediction performance in different scenarios;

[0057] As shown in Table 1, under three different propagation scenarios, the proposed prediction model outperforms the comparative models in all evaluation metrics, including mean square error, root mean square error, mean absolute error, and coefficient of determination. Compared to BPNN and LSTM models, the proposed method shows a significant downward trend in error metrics, and the coefficient of determination is closer to 1, indicating a higher consistency between the prediction results and the actual channel state. The BPNN model, due to its failure to explicitly model the temporal correlation of the channel, exhibits significant deviations in scenarios with drastic multipath power changes. While the LSTM model possesses some temporal modeling capabilities, its ability to characterize high-frequency changes and complex multipath structures in the channel is limited, thus restricting its prediction accuracy. In contrast, the prediction model proposed in this embodiment can more accurately reconstruct the overall distribution and local details of the channel power delay spectrum, demonstrating 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, achieving high-precision prediction of channel state under multiple scenario conditions, demonstrating good generalization ability and robustness.

[0058] Example 2 Based on Embodiment 1, this embodiment provides a UAV ground channel prediction system based on adaptive Fourier features, including: 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. The context information extraction module is configured to extract the context information of the delay of signals from different paths in multipath propagation based on the channel impulse response and delay AFF mapping definition of the channel signal of the previous frame. The timing coding module is configured to model the correlation between different multipath signals based on the time-delay AFF mapping, construct the cross-path kernel, mix the historical channel impulse response sequence, obtain a new cross-path channel representation, input it to the GRU timing encoder, and extract the time context code. 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.

[0059] Furthermore, the timing coding module is configured to perform the following steps: Stack the delay AFF mapping features of all multipaths into a matrix to obtain the delay AFF dictionary; 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. 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.

[0060] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0061] Example 3 Based on Embodiment 1, this embodiment provides a UAV ground channel prediction system based on adaptive Fourier features, including: a data acquisition device and a processor; Acquisition device, used to acquire historical channel impulse response sequences; The processor is configured to execute the steps of the UAV ground channel prediction method based on adaptive Fourier features as described in Example 1.

[0062] Specifically, the channel measurement system used by the acquisition device includes an air signal transmitting unit and a ground receiving unit.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0064] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

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 respectively; 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 contextual information of the delay of signals of different paths in multipath propagation. Based on the time-delay AFF mapping model, the correlation between different multipath signals is modeled, the cross-path kernel is constructed and the historical channel impulse response sequence is mixed to obtain a new cross-path channel representation, which is then input into the GRU time encoder to extract the time context code. 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: The method for extracting the timestamp and multipath delay corresponding to each historical channel impulse response, and performing adaptive Fourier feature mapping to obtain the time AFF mapping and delay AFF mapping includes 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 of different paths in multipath propagation, resulting in a time delay AFF mapping.

3. The UAV-to-ground channel prediction method based on adaptive Fourier features as described in claim 1, characterized in that: Based on the time-delay AFF mapping model, the correlation between different multipath signals is modeled, and a cross-path kernel is constructed to mix historical channel impulse response sequences to obtain a new cross-path channel representation, including 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. A dictionary representing delay AFFs, 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.

4. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 1, characterized in that: The delay AFF mapping features of all multipath propagation paths are stacked into a matrix to obtain the delay AFF dictionary. Based on the channel impulse response of the previous frame's channel signal and the delay AFF dictionary, 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, A dictionary representing delay AFFs, This represents the channel impulse response of the channel signal in the previous frame; This represents the base of the delay AFF.

5. The UAV-to-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.

6. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 5, 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.

7. The UAV ground channel prediction method based on adaptive Fourier features as described in claim 6, 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. The channel impulse response samples are used as input to distinguish between real and fake channel data.

8. A UAV 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. The context information extraction module is configured to extract the context information of the delay of signals from different paths in multipath propagation based on the channel impulse response and delay AFF mapping definition of the channel signal of the previous frame. The timing coding module is configured to model the correlation between different multipath signals based on the time-delay AFF mapping, construct the cross-path kernel, mix the historical channel impulse response sequence, obtain a new cross-path channel representation, input it to the GRU timing encoder, and extract the time context code. 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.

9. The UAV ground channel prediction system based on adaptive Fourier features as described in claim 8, characterized in that, The timing coding module is configured to perform 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. A dictionary representing delay AFFs, Indicates transpose; 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.

10. A UAV 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 1-7.

Images