A digital twin channel prediction method for high mobility MIMO

Abstract

The application discloses a digital twin channel prediction method for high-mobility MIMO, comprising: constructing a channel evolution model and determining a channel aging interval; obtaining uniform sparse pilot observation data; transmitting the uniform sparse pilot observation data from a physical channel layer to a digital twin layer and establishing a sliding time window; using the uniform sparse pilot observation data in the sliding time window to perform continuous channel trajectory reconstruction and obtain a continuous historical channel trajectory; predicting a future channel state of the continuous historical channel trajectory; periodically updating the sliding time window in the digital twin layer according to a preset pilot detection interval, and cyclically performing continuous historical channel trajectory reconstruction, future channel state prediction and sliding time window updating. The application solves the problems of inconsistent distribution of sparse pilot windows, multi-step extrapolation error accumulation, prediction phase deviation destroying beam orthogonality and background update blocking online reasoning in a high-mobility non-stationary channel.

Description

Technical Field

[0001] This invention belongs to the field of channel state information prediction technology, specifically relating to a digital twin channel prediction method for high mobility MIMO; the method of this invention is applicable to high-speed railway communication, vehicle-to-everything communication, and other multi-antenna communication scenarios with significant Doppler frequency shift and integrated processing delay. Background Technology

[0002] Massive Multiple-Input Multiple-Output (MMIMO) technology forms highly directional beams through a large array of antennas, which can make full use of spatial degrees of freedom to improve spectral efficiency and system capacity. The above gains depend on the base station obtaining accurate and timely Channel State Information (CSI) during downlink transmission.

[0003] In highly mobile scenarios such as high-speed rail and vehicle-to-everything communication, terminal movement generates significant Doppler frequency shift and rapid time-varying fading. After obtaining pilot observations, base stations still need to perform channel estimation, scheduling, precoding calculations, and hardware pipeline processing. The channel state information used in actual downlink transmission may have deviated from the instantaneous true channel, resulting in channel state information aging. Channel state information aging will destroy the orthogonality between multiple spatial streams, causing inter-stream crosstalk, beam mismatch, and system capacity reduction.

[0004] To reduce pilot overhead, existing schemes often employ sparse or non-uniform pilots and utilize autoregressive models, recurrent neural networks, or Transformers to predict future channel state information. However, non-uniform pilots are prone to losing the translation invariance of sampling positions when the prediction window slides continuously, leading to inconsistent observation distributions between different windows. When using uniform sparse pilots, existing time-series models struggle to recover rapid channel changes. Meanwhile, recursive multi-step prediction uses the prediction result from the previous time step as the input for the next, which can easily cause error accumulation in fast, non-stationary channels. Using only the normalized mean square error as the training objective also fails to directly reflect the impact of complex channel phase deviations on singular value decomposition beamforming and multi-stream orthogonality.

[0005] Furthermore, real-time communication requires predictive inference to be completed within a strict baseband latency budget, while model fine-tuning involves backpropagation and parameter updates, the execution latency of which is typically greater than that of a single forward inference and varies with hardware load. If online inference and model updates are executed serially, the communication link will be blocked by background training; if the model is not updated for a long time, it will be difficult to adapt to non-stationary environments where speed, scatterers, and channel statistical characteristics change over time.

[0006] Therefore, a channel prediction technique is needed that can take into account sparse pilot reconstruction, continuous multi-step prediction, physical phase fidelity, and asynchronous online deployment. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the present invention aims to provide a digital twin channel prediction method for high-mobility MIMO, thereby solving the problems of inconsistent sparse pilot window distribution, multi-step extrapolation error accumulation, prediction phase deviation disrupting beam orthogonality, and background update blocking online inference in high-mobility non-stationary channels.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The present invention provides a digital twin channel prediction method for high mobile MIMO, comprising the following steps:

[0010] S1: Construct a channel evolution model for high-mobility multi-antenna communication scenarios, determine the channel aging interval based on the mobile terminal's motion state, Doppler frequency shift, and comprehensive processing delay; acquire uniform sparse pilot observation data transmitted by the mobile terminal and received by the base station;

[0011] S2: Construct a physical channel layer and a digital twin layer. Transmit the uniform sparse pilot observation data obtained in step S1 from the physical channel layer to the digital twin layer, and establish a sliding time window in the digital twin layer corresponding to the current physical channel layer.

[0012] S3: Continuous channel trajectory reconstruction is performed on the uniform sparse pilot observation data within the sliding time window in the digital twin layer using a continuous channel trajectory reconstruction network to obtain continuous historical channel trajectories;

[0013] S4: Use a future channel state prediction network to predict the future channel state of continuous historical channel trajectories;

[0014] S5: The sliding time window in the digital twin layer is periodically updated according to the preset pilot detection interval, and the continuous historical channel trajectory reconstruction, future channel state prediction and sliding time window update are performed in a loop.

[0015] Furthermore, the method also includes:

[0016] S6: Determine the prediction error based on the future channel state prediction results and new pilot observation data, and construct a joint optimization loss function that simultaneously constrains the channel numerical prediction accuracy and spatial phase consistency based on the prediction error;

[0017] S7: Based on the joint optimization loss function obtained in step S6, keep the network parameters of the continuous channel trajectory reconstruction network unchanged, and asynchronously update the network parameters of the future channel state prediction network; during the parameter update, continue to use the network parameters before the update for future channel state prediction, and after the parameter update is completed, use the updated network parameters for subsequent future channel state prediction.

[0018] Further, step S1 specifically includes:

[0019] A channel evolution model is constructed to describe the correlation between channel states at adjacent time points in high-mobility multi-antenna communication scenarios. A first-order Gaussian-Markov relation is used to correlate the channel states at adjacent time points, expressed as:

[0020] ;

[0021] in, Indicates time The complex downlink channel matrix, and ; Indicates time The complex downlink channel matrix; Represents the set of complex numbers; Indicates the number of transmitting antennas at the base station; Indicates the number of receiving antennas in the mobile terminal; This represents the total time interval between the pilot observation time and the actual downlink transmission time; Indicates the maximum Doppler frequency shift and time interval The relevant channel time correlation coefficient is used to characterize the degree of correlation between the channel states at two different times; Indicates time Relative to time Channel aging variation term;

[0022] Determine the channel aging interval between the pilot observation time and the actual downlink transmission time, and express the total time interval corresponding to the channel aging interval as:

[0023] ;

[0024] in, The pilot detection interval represents a uniformly sparse pilot, i.e., the time interval between two adjacent pilot observations; This represents the overall processing delay generated from obtaining pilot observation data to executing actual downlink transmission. The overall processing delay includes the delays generated by channel estimation, channel prediction, scheduling, precoding calculation, and baseband hardware pipeline processing.

[0025] The maximum Doppler frequency shift is determined based on the motion state of the mobile terminal, and combined with the total time interval. Determine the degree of channel time correlation;

[0026] Acquiring uniformly sparse pilot observation data sent by the mobile terminal and received by the base station, specifically: the mobile terminal follows a preset pilot detection interval. A pilot signal is transmitted, the base station receives the pilot signal, and obtains the corresponding complex channel state observation data according to the pilot signal received at each pilot observation time; a uniform sparse pilot observation sequence is formed in the time dimension according to the pilot detection interval.

[0027] Further, step S2 specifically includes:

[0028] A physical channel layer is constructed, and uniform sparse pilot observation data is transmitted to the digital twin layer as the observation basis for the digital twin layer to characterize the current state of the physical channel layer.

[0029] The physical channel layer includes: a base station multi-antenna array, a mobile terminal, a wireless propagation channel, and pilot signals and data signals transmitted between the base station and the mobile terminal; the uniform sparse pilot observation data received by the base station is used to characterize the physical channel state at the corresponding pilot observation time.

[0030] A digital twin layer is constructed, and a sliding time window corresponding to the current physical channel layer is established in the digital twin layer. The uniform sparse pilot observation data obtained in step S1 is stored in the sliding time window according to the pilot observation time.

[0031] Furthermore, step S3 specifically includes:

[0032] A continuous channel trajectory reconstruction network is constructed. The continuous channel trajectory reconstruction network is the first-level channel reconstruction network in a two-level cascaded network. It is used to reconstruct the channel state of the unobserved time position within the sliding time window based on the uniform sparse pilot observation data and the corresponding pilot position information within the sliding time window, so as to obtain the continuous historical channel trajectory.

[0033] The complex channel state information at each pilot observation time is separated into real and imaginary parts and arranged along the feature dimension to form a sparse pilot input tensor. Construct a binary pilot position mask based on the pilot positions corresponding to the uniform sparse pilot observation data. In locations where actual pilot observations exist, the corresponding mask element is set to 1; in locations where no actual pilot observations exist, the corresponding mask element is set to 0.

[0034] Input the sparse pilot into the tensor With the binary pilot position mask The splicing is performed, and the splicing result is linearly projected to map the splicing result to the latent feature space; absolute position encoding is added to the features obtained by linear projection to preserve the chronological order of each time sampling position within the sliding time window;

[0035] The features after adding absolute position encoding are subjected to concatenated one-dimensional residual dilated convolution processing to extract local multi-scale channel change features. The one-dimensional residual dilated convolution unit adopts different dilation rates to expand the temporal receptive field. Among them, the small dilation rate is used to extract fast fading and local phase change features, and the large dilation rate is used to extract long-term channel evolution features. Each one-dimensional residual dilated convolution unit sequentially performs one-dimensional dilated convolution, batch normalization, Gaussian error linear unit activation, and residual connection.

[0036] The features processed by cascaded one-dimensional residual dilated convolution are input into a time-domain encoder based on a self-attention mechanism to establish long-range dependencies between different time positions within a sliding time window; then, output projection is performed to obtain the original continuous channel reconstruction result. ;

[0037] The original continuous channel reconstruction result Perform mask fusion, retain the corresponding physical observation values ​​at the actual pilot observation locations, and use the reconstructed values ​​output by the first-level channel reconstruction network at the unobserved locations to obtain continuous historical channel trajectories, represented as:

[0038] ;

[0039] in, This represents the continuous historical channel trajectory obtained after mask fusion; Represents the binary pilot position mask Tensors with the same dimensions and all elements being 1; This represents the element-wise multiplication operation between elements at corresponding positions; This represents the inverse mask corresponding to the unobserved position, where the unobserved position is set to 1 and the actual pilot observation position is set to 0.

[0040] Furthermore, step S4 specifically includes:

[0041] A future channel state prediction network is constructed. This network serves as the second-level channel prediction network in a two-stage cascaded network. Based on the continuous historical channel trajectories obtained in step S3, it extracts channel evolution features at different time scales and spatial correlation features between different antenna links. It then predicts the channel state at multiple future moments within a forward prediction interval corresponding to a pilot detection interval in a non-autoregressive manner. The time domain length of the forward prediction interval is the same as the pilot detection interval, and the prediction result includes the channel state corresponding to the next pilot observation moment. The continuous historical channel trajectories obtained in step S3 are arranged according to different antenna links, so that the historical channel state sequence of each antenna link within the sliding time window is input as an independent one-dimensional time series into the second-level channel prediction network.

[0042] The historical channel state sequences of each antenna link are subjected to dual-path embedding processing of multi-scale convolution and linear mapping. The multi-scale convolution branch uses multiple one-dimensional convolution kernels of different scales to extract channel evolution features at different time scales. Small-scale one-dimensional convolution kernels are used to extract short-term rapid fading and phase change features, medium-scale one-dimensional convolution kernels are used to extract medium-scale change trends, and large-scale one-dimensional convolution kernels are used to extract slowly varying envelope features. The linear mapping branch performs linear mapping on the complete historical channel state sequence to preserve the overall channel evolution trend.

[0043] By adaptively fusing the outputs of the two branches using learnable fusion parameters, a fused embedding feature is obtained. , is represented as:

[0044] ;

[0045] in, This represents the local multi-scale channel characteristics of the multi-scale convolution branch output; This represents the overall channel evolution characteristics of the linear mapping branch output; This represents the fused embedding feature obtained after adaptive fusion of the two branches; Represents the learnable fusion parameters; This represents a logistic activation function used to fuse the parameters. Mapped to fusion weights between 0 and 1; This represents the fusion weights corresponding to the linear mapping branches;

[0046] The fused embedding features of each antenna link within the entire sliding time window are treated as a variable representation unit, and cross-variable self-attention processing is performed between different antenna links to utilize the spatial correlation between different antenna links in a co-propagation environment; for the ... The antenna link and the first For each antenna link, the attention weights are expressed as follows:

[0047] ;

[0048] The value vectors of different antenna links are weighted and aggregated according to the attention weights to obtain the first... The spatial compensation characteristics of an antenna link are expressed as follows:

[0049] ;

[0050] in, Indicates the first The antenna link pair of the first Attention weights for each antenna link; Indicates the first The query vector for each antenna link; Indicates the first The key vector of each antenna link; Indicates the first The value vector of each antenna link; In the process of summation, the first... The key vector of each antenna link; Indicates the transpose operation; This represents the feature dimension of the key vector; This indicates the number of antenna links participating in the intervariate self-attention process; Indicates the summation index; Represents an exponential function; Indicates the first Spatial compensation features obtained by transvariable self-attention processing of individual antenna links;

[0051] The features processed by intervariate self-attention are output-mapped, and the final residuals are connected to generate the future channel state prediction results, as follows:

[0052] ;

[0053] in, This indicates the channel state at the end of the sliding time window of the continuous historical channel trajectory; Indicates the length of the historical channel state sequence within the sliding time window; This represents the future channel change predicted by the second-level channel prediction network based on historical channel evolution patterns. This indicates the prediction result of the future channel state.

[0054] Further, step S5 specifically includes:

[0055] The sliding time window is moved forward by one pilot detection interval along the time direction. The earliest historical pilot observation data in the sliding time window is removed. The new pilot observation data and its corresponding pilot position information are stored in the sliding time window, and the binary pilot position mask is updated simultaneously, so that the updated sliding time window continues to maintain a uniform pilot sampling structure.

[0056] Based on the updated sliding time window, step S3 is executed again, and the continuous channel trajectory is reconstructed on the updated pilot observation data using the continuous channel trajectory reconstruction network to obtain the updated continuous historical channel trajectory; then, based on the updated continuous historical channel trajectory, step S4 is executed again, and the updated future channel state prediction result is obtained using the future channel state prediction network.

[0057] During each sliding time window update, the new pilot observation data is used as the new physical observation anchor point, so that the continuous historical channel trajectory and the prediction results of future channel state are periodically constrained by the latest physical channel observation. This reduces the accumulation of prediction errors caused by long-term open-loop extrapolation and keeps the continuous channel reconstruction and future channel state prediction process within a wide stable range corresponding to the current channel statistical characteristics.

[0058] Further, step S6 specifically includes:

[0059] The prediction channel matrix corresponding to the future channel state prediction results obtained in step S4 Perform singular value decomposition, represented as:

[0060] ;

[0061] in, This represents the unitary matrix composed of the left singular vectors of the prediction channel matrix; This represents a matrix composed of the singular values ​​of the prediction channel matrix; This represents a unitary matrix composed of the right singular vectors of the prediction channel matrix; the base station uses this matrix... Generate a downlink precoding matrix, which the mobile terminal then utilizes. Perform receive merging to obtain the receive merge matrix;

[0062] When new pilot observation data acquired periodically in step S5 arrives, and its pilot observation time corresponds to the corresponding prediction time in the previous future channel state prediction results, the physical channel observation reference state at the corresponding time is obtained based on the new pilot observation data, and the prediction error matrix between the observation reference channel matrix and the prediction channel matrix corresponding to the physical channel observation reference state is calculated, as follows:

[0063] ;

[0064] in, Represents the prediction error matrix; This represents the observation reference channel matrix corresponding to the time of the new pilot observation data; This represents the prediction channel matrix at the corresponding time point;

[0065] Applying the precoding matrix and receive-combining matrix obtained based on the predicted channel matrix to the real physical channel yields the actual equivalent baseband signal, expressed as:

[0066] ;

[0067] in, This represents the equivalent baseband received signal vector after merging; Indicates the downlink transmitted symbol vector; Represents the noise vector of the mobile terminal; This represents the desired signal component formed after spatial flow decoupling based on the predicted channel matrix; Represents the prediction error matrix This results in residual inter-flow crosstalk components in the space;

[0068] A joint optimization loss function is constructed based on the prediction error, which simultaneously constrains the channel numerical prediction error and the phase direction deviation of the complex channel in vector space, and is expressed as:

[0069] ;

[0070] in, Represents the joint optimization loss function; This represents the mean squared error loss, used to measure the numerical difference between the actual channel state and the predicted channel state. It represents the generalized cosine phase consistency loss, which is used to measure the directional difference between the actual channel state and the predicted channel state in the complex vector space; Indicates the loss weight, and This is used to balance numerical prediction accuracy with spatial phase consistency.

[0071] The real channel matrix and prediction channel matrix Vectorized into real channel vectors respectively and predicted channel vector The generalized cosine phase consistency loss Represented as:

[0072] ;

[0073] in, Represents the true channel vector The conjugate transpose of; Represents the magnitude of the complex inner product of the true channel vector and the predicted channel vector; Represents the true channel vector The second norm; Represents the predicted channel vector The second norm; This indicates that the expected value or average value of the sample is calculated.

[0074] Further, step S7 specifically includes:

[0075] Keeping the network parameters of the continuous channel trajectory reconstruction network unchanged, the gradient of the joint optimization loss function with respect to the network parameters of the future channel state prediction network is calculated, and the network parameters of the future channel state prediction network are updated asynchronously based on the gradient; the updated network parameters of the future channel state prediction network are expressed as:

[0076] ;

[0077] in, This represents the network parameters of the future channel state prediction network before the asynchronous update begins; Indicates the delay after asynchronous processing. Complete the updated network parameters of the future channel state prediction network; Indicates the start time of the asynchronous update; This indicates the processing delay experienced by the asynchronous update from start to completion; The learning rate represents the network parameter update rate; This represents the joint optimization loss function calculated based on the network parameters of the network predicting the current future channel state. This represents the joint optimization loss function relative to the network parameters of the future channel state prediction network. The gradient;

[0078] During the asynchronous parameter update of the future channel state prediction network, the network parameters before the update are still used to perform future channel state prediction.

[0079] The future channel state prediction network with updated network parameters is used to predict the subsequent future channel state.

[0080] The beneficial effects of this invention are:

[0081] 1. This invention decouples sparse pilot reconstruction and future CSI prediction into two cascaded stages, avoiding optimization conflicts caused by a single network simultaneously undertaking missing value recovery and multi-step extrapolation.

[0082] 2. This invention utilizes mask fusion to preserve measured pilot signals, ensuring that the digital twin channel trajectory is always anchored by physical observations, thereby reducing reconstruction drift.

[0083] 3. This invention utilizes dual-stream cross-scale embedding and cross-variable self-attention to simultaneously capture fast phase rotation, slow fading envelope, and antenna link spatial correlation, which helps to improve prediction robustness in non-stationary scenarios and improves downlink beamforming performance overall under medium-to-high signal-to-noise ratio conditions.

[0084] 4. This invention utilizes a dynamic sliding time window to periodically inject real pilot signals, reducing error propagation in recursive multi-step prediction and meeting the requirements for continuous communication.

[0085] 5. In this invention, online inference and background training are asynchronously decoupled, so that the downlink is not interrupted due to backpropagation delay, and it is adapted to the actual computing power and scheduling fluctuations of the baseband processor.

[0086] 6. The present invention combines the optimization of numerical accuracy and complex phase consistency, which can reduce the off-diagonal crosstalk caused by prediction errors in the SVD equivalent channel and improve the orthogonality of prediction beamforming and system capacity. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the method of the present invention.

[0088] Figure 2 This is a normalized mean square error plot showing the variation of signal-to-noise ratio for different prediction schemes.

[0089] Figure 3 The system capacity diagram is shown for downlink beamforming based on the prediction channel matrix obtained from each prediction scheme under different signal-to-noise ratios. Detailed Implementation

[0090] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0091] Reference Figure 1 As shown, the present invention provides a digital twin channel prediction method for high mobile MIMO, comprising the following steps:

[0092] S1: Construct a channel evolution model for high-mobility multi-antenna communication scenarios, and determine the channel aging interval based on the mobile terminal's motion state, Doppler frequency shift, and comprehensive processing delay; acquire uniform sparse pilot observation data transmitted by the mobile terminal and received by the base station; specifically including:

[0093] A channel evolution model is constructed to describe the correlation between channel states at adjacent time points in high-mobility multi-antenna communication scenarios. A first-order Gaussian-Markov relation is used to correlate the channel states at adjacent time points, expressed as:

[0094] ;

[0095] in, Indicates time The complex downlink channel matrix, and ; Indicates time The complex downlink channel matrix; Represents the set of complex numbers; Indicates the number of transmitting antennas at the base station; Indicates the number of receiving antennas in the mobile terminal; This represents the total time interval between the pilot observation time and the actual downlink transmission time; Indicates the maximum Doppler frequency shift and time interval The relevant channel time correlation coefficient is used to characterize the degree of correlation between the channel states at two different times; Indicates time Relative to time Channel aging variation term;

[0096] The channel evolution model corresponds to having Base stations with one transmitting antenna and having A downlink wireless channel between mobile terminals with multiple receiving antennas; the movement of the mobile terminals causes a Doppler frequency shift in the wireless propagation path, resulting in changes in the channel state at different times; the channel time correlation coefficient. Used to characterize the correlation between channel states before and after time changes;

[0097] Determine the channel aging interval between the pilot observation time and the actual downlink transmission time, and express the total time interval corresponding to the channel aging interval as:

[0098] ;

[0099] in, The pilot detection interval represents a uniformly sparse pilot, i.e., the time interval between two adjacent pilot observations; This represents the overall processing delay generated from obtaining pilot observation data to executing actual downlink transmission. The overall processing delay includes the delays generated by channel estimation, channel prediction, scheduling, precoding calculation, and baseband hardware pipeline processing.

[0100] The maximum Doppler frequency shift is determined based on the motion state of the mobile terminal, and combined with the total time interval. Determine the degree of channel time correlation; as the maximum Doppler frequency shift increases or the total time interval increases... The correlation between the channel state at the pilot observation time and the channel state at the actual downlink transmission time decreases, thereby increasing the channel aging process.

[0101] Acquiring uniformly sparse pilot observation data sent by the mobile terminal and received by the base station, specifically: the mobile terminal follows a preset pilot detection interval. A pilot signal is transmitted, the base station receives the pilot signal, and obtains the corresponding complex channel state observation data according to the pilot signal received at each pilot observation time; a uniform sparse pilot observation sequence is formed in the time dimension according to the pilot detection interval.

[0102] It should be noted that the uniform sparse pilot observation data and its corresponding pilot position information are used as inputs to the subsequent channel reconstruction and prediction process, so as to reconstruct the continuous historical channel trajectory based on the uniform sparse pilot observation data, and predict the channel state information at future times within the channel aging interval based on the continuous historical channel trajectory.

[0103] S2: Construct a physical channel layer and a digital twin layer. Transmit the uniform sparse pilot observation data obtained in step S1 from the physical channel layer to the digital twin layer, and establish a sliding time window in the digital twin layer corresponding to the current physical channel layer; specifically including:

[0104] Construct a physical channel layer and transmit the uniform sparse pilot observation data from step S1 to the digital twin layer as the observation basis for the digital twin layer to characterize the current state of the physical channel layer.

[0105] The physical channel layer includes: a base station multi-antenna array, a mobile terminal, a wireless propagation channel, and pilot signals and data signals transmitted between the base station and the mobile terminal; the uniform sparse pilot observation data received by the base station is used to characterize the physical channel state at the corresponding pilot observation time.

[0106] A digital twin layer is constructed, and a sliding time window corresponding to the current physical channel layer is established in the digital twin layer. The uniform sparse pilot observation data obtained in step S1 is stored in the sliding time window according to the pilot observation time.

[0107] The uniform sparse pilots are distributed according to a fixed pilot detection interval in the time dimension, so that the sliding time window maintains a consistent pilot sampling structure in subsequent updates. This reduces the changes in input data distribution caused by the change of pilot sampling position with the sliding time window, and provides structurally consistent input data for the subsequent first-level channel reconstruction network to reconstruct continuous historical channel trajectories.

[0108] S3: Utilize a continuous channel trajectory reconstruction network to reconstruct the continuous historical channel trajectory from the uniformly sparse pilot observation data within the sliding time window of the digital twin layer; specifically including:

[0109] A continuous channel trajectory reconstruction network is constructed. The continuous channel trajectory reconstruction network is the first-level channel reconstruction network in a two-level cascaded network. It includes an input coding module, a local multi-scale feature extraction module, a time domain coding module, an output projection module, and a mask fusion module. It is used to reconstruct the channel state of the unobserved time position within the sliding time window based on the uniform sparse pilot observation data and the corresponding pilot position information within the sliding time window, so as to obtain the continuous historical channel trajectory.

[0110] The complex channel state information at each pilot observation time is separated into real and imaginary parts and arranged along the feature dimension to form a sparse pilot input tensor. Construct a binary pilot position mask based on the pilot positions corresponding to the uniform sparse pilot observation data. In locations where actual pilot observations exist, the corresponding mask element is set to 1; in locations where no actual pilot observations exist, the corresponding mask element is set to 0.

[0111] Input the sparse pilot into the tensor With the binary pilot position mask The splicing is performed, and the splicing result is linearly projected to map the splicing result to the latent feature space; absolute position encoding is added to the features obtained by linear projection to preserve the chronological order of each time sampling position within the sliding time window;

[0112] The features after adding absolute position encoding are subjected to concatenated one-dimensional residual dilated convolution processing to extract local multi-scale channel change features. The one-dimensional residual dilated convolution unit adopts different dilation rates to expand the temporal receptive field. Among them, a small dilation rate is used to extract fast fading and local phase change features, and a large dilation rate is used to extract long-term channel evolution features. Each one-dimensional residual dilated convolution unit sequentially performs one-dimensional dilated convolution, batch normalization, Gaussian error linear unit activation, and residual connection. In one embodiment, the dilation rates are set to 1, 2, and 4 in sequence.

[0113] The features processed by cascaded one-dimensional residual dilated convolution are input into a time-domain encoder based on a self-attention mechanism to establish long-range dependencies between different time positions within a sliding time window; then, output projection is performed to obtain the original continuous channel reconstruction result. ;

[0114] The original continuous channel reconstruction result Perform mask fusion, retain the corresponding physical observation values ​​at the actual pilot observation locations, and use the reconstructed values ​​output by the first-level channel reconstruction network at the unobserved locations to obtain continuous historical channel trajectories, represented as:

[0115] ;

[0116] in, This represents the continuous historical channel trajectory obtained after mask fusion; Represents the binary pilot position mask Tensors with the same dimensions and all elements being 1; This represents the element-wise multiplication operation between elements at corresponding positions; This represents the inverse mask corresponding to the unobserved position, where the unobserved position is set to 1 and the actual pilot observation position is set to 0.

[0117] Specifically, through mask fusion, the sparse pilot input tensor is used at the actual pilot observation location. The physical observations in the original continuous channel reconstruction results are replaced. The corresponding network reconstruction values ​​in the data are used to retain the original continuous channel reconstruction results at unobserved locations. This ensures that the continuous historical channel trajectory is constrained by real physical observations at the actual pilot observation locations, and completes the channel state at unobserved time locations.

[0118] S4: Use a future channel state prediction network to predict the future channel state of continuous historical channel trajectories;

[0119] The prediction time domain length of the future channel state prediction is the same as the pilot detection interval in step S1, and the prediction target must include the channel state corresponding to the next pilot observation time.

[0120] And, step S4 specifically includes:

[0121] A future channel state prediction network is constructed. This network, as the second-level channel prediction network in a two-level cascaded network, includes a dual-path embedding module, an adaptive fusion module, a cross-variable self-attention module, and a prediction output module. Based on the continuous historical channel trajectories obtained in step S3, it extracts channel evolution features at different time scales and spatial correlation features between different antenna links. It then predicts the channel state at multiple future moments within a forward prediction interval corresponding to a future pilot detection interval in a non-autoregressive manner. The time domain length of the forward prediction interval is the same as the pilot detection interval, and the prediction result includes the channel state corresponding to the next pilot observation moment. The continuous historical channel trajectories obtained in step S3 are arranged according to different antenna links, so that the historical channel state sequence of each antenna link within a sliding time window is input as an independent one-dimensional time series to the second-level channel prediction network.

[0122] The historical channel state sequences of each antenna link are subjected to dual-path embedding processing of multi-scale convolution and linear mapping. The multi-scale convolution branch uses multiple one-dimensional convolution kernels of different scales to extract channel evolution features at different time scales. Small-scale one-dimensional convolution kernels are used to extract short-term rapid fading and phase change features, medium-scale one-dimensional convolution kernels are used to extract medium-scale change trends, and large-scale one-dimensional convolution kernels are used to extract slowly varying envelope features. The linear mapping branch performs linear mapping on the complete historical channel state sequence to preserve the overall channel evolution trend. In one embodiment, the sizes of the multiple one-dimensional convolution kernels are set to 3, 11, and 19, respectively.

[0123] By adaptively fusing the outputs of the two branches using learnable fusion parameters, a fused embedding feature is obtained. , is represented as:

[0124] ;

[0125] in, This represents the local multi-scale channel characteristics of the multi-scale convolution branch output; This represents the overall channel evolution characteristics of the linear mapping branch output; This represents the fused embedding feature obtained after adaptive fusion of the two branches; Represents the learnable fusion parameters; This represents a logistic activation function used to fuse the parameters. Mapped to fusion weights between 0 and 1; This represents the fusion weights corresponding to the linear mapping branches;

[0126] The fused embedding features of each antenna link within the entire sliding time window are treated as a variable representation unit, and cross-variable self-attention processing is performed between different antenna links to utilize the spatial correlation between different antenna links in a co-propagation environment; for the ... The antenna link and the first For each antenna link, the attention weights are expressed as follows:

[0127] ;

[0128] The value vectors of different antenna links are weighted and aggregated according to the attention weights to obtain the first... The spatial compensation characteristics of an antenna link are expressed as follows:

[0129] ;

[0130] in, Indicates the first The antenna link pair of the first Attention weights for each antenna link; Indicates the first The query vector for each antenna link; Indicates the first The key vector of each antenna link; Indicates the first The value vector of each antenna link; In the process of summation, the first... The key vector of each antenna link; Indicates the transpose operation; This represents the feature dimension of the key vector; This indicates the number of antenna links participating in the intervariate self-attention process; Indicates the summation index; Represents an exponential function; Indicates the first Spatial compensation features obtained by transvariable self-attention processing of individual antenna links;

[0131] The features processed by intervariate self-attention are output-mapped, and the final residuals are connected to generate the future channel state prediction results, as follows:

[0132] ;

[0133] in, This indicates the channel state at the end of the sliding time window of the continuous historical channel trajectory; Indicates the length of the historical channel state sequence within the sliding time window; This represents the future channel change predicted by the second-level channel prediction network based on historical channel evolution patterns. This indicates the prediction result of the future channel state.

[0134] For a prediction interval containing multiple future times, the second-level channel prediction network outputs the channel change corresponding to each future time in parallel, and compares it with the terminal channel state. By performing residual superposition, the prediction results of the previous future time are not used as the input of the next future time, thereby reducing the error accumulation caused by time-by-time recursive prediction.

[0135] S5: Periodically update the sliding time window in the digital twin layer according to the preset pilot detection interval, and cyclically perform continuous historical channel trajectory reconstruction, future channel state prediction, and sliding time window update; specifically including:

[0136] New pilot observation data are periodically acquired according to the pilot detection interval determined in step S1, and the arrival of new pilot observation data is used as the trigger condition for updating the sliding time window.

[0137] The sliding time window is moved forward by one pilot detection interval along the time direction. The earliest historical pilot observation data in the sliding time window is removed. The new pilot observation data and its corresponding pilot position information are stored in the sliding time window, and the binary pilot position mask is updated simultaneously, so that the updated sliding time window continues to maintain a uniform pilot sampling structure.

[0138] Based on the updated sliding time window, step S3 is executed again, and the first-level channel reconstruction network is used to reconstruct the continuous channel trajectory of the updated uniform sparse pilot observation data to obtain the updated continuous historical channel trajectory; then, based on the updated continuous historical channel trajectory, step S4 is executed again, and the second-level channel prediction network is used to obtain the updated future channel state prediction result.

[0139] During each sliding time window update, the new pilot observation data is used as the new physical observation anchor point, so that the continuous historical channel trajectory and the prediction results of future channel state are periodically constrained by the latest physical channel observation. This reduces the accumulation of prediction errors caused by long-term open-loop extrapolation and keeps the continuous channel reconstruction and future channel state prediction process within a wide stable range corresponding to the current channel statistical characteristics.

[0140] In a preferred embodiment, the method of the present invention further includes:

[0141] S6: Determine the prediction error based on the future channel state prediction results and new pilot observation data, and construct a joint optimization loss function that simultaneously constrains the channel numerical prediction accuracy and spatial phase consistency based on the prediction error;

[0142] S7: Based on the joint optimization loss function obtained in step S6, keep the network parameters of the continuous channel trajectory reconstruction network unchanged, and asynchronously update the network parameters of the future channel state prediction network; during the parameter update, continue to use the network parameters before the update for future channel state prediction, and after the parameter update is completed, use the updated network parameters for subsequent future channel state prediction.

[0143] More specifically, step S6 includes:

[0144] The prediction channel matrix corresponding to the future channel state prediction results obtained in step S4 Perform singular value decomposition, represented as:

[0145] ;

[0146] in, This represents the unitary matrix composed of the left singular vectors of the prediction channel matrix; This represents a matrix composed of the singular values ​​of the prediction channel matrix; This represents a unitary matrix composed of the right singular vectors of the prediction channel matrix; the base station uses this matrix... Generate a downlink precoding matrix, which the mobile terminal then utilizes. Perform receive merging to obtain the receive merge matrix;

[0147] When new pilot observation data acquired periodically in step S5 arrives, and its pilot observation time corresponds to the corresponding prediction time in the previous future channel state prediction results, the physical channel observation reference state at the corresponding time is obtained based on the new pilot observation data, and the prediction error matrix between the observation reference channel matrix and the prediction channel matrix corresponding to the physical channel observation reference state is calculated, as follows:

[0148] ;

[0149] in, Represents the prediction error matrix; This represents the observation reference channel matrix corresponding to the time of the new pilot observation data; This represents the prediction channel matrix at the corresponding time point;

[0150] Applying the precoding matrix and receive-combining matrix obtained based on the predicted channel matrix to the real physical channel yields the actual equivalent baseband signal, expressed as:

[0151] ;

[0152] in, This represents the equivalent baseband received signal vector after merging; Indicates the downlink transmitted symbol vector; Represents the noise vector of the mobile terminal; This represents the desired signal component formed after spatial flow decoupling based on the predicted channel matrix; Represents the prediction error matrix This results in residual inter-flow crosstalk components in the space;

[0153] The off-diagonal component of the residual spatial flow crosstalk component reflects the destruction of spatial flow orthogonality caused by prediction error.

[0154] A joint optimization loss function is constructed based on the prediction error, which simultaneously constrains the channel numerical prediction error and the phase direction deviation of the complex channel in vector space, and is expressed as:

[0155] ;

[0156] in, Represents the joint optimization loss function; This represents the mean squared error loss, used to measure the numerical difference between the actual channel state and the predicted channel state. It represents the generalized cosine phase consistency loss, which is used to measure the directional difference between the actual channel state and the predicted channel state in the complex vector space; Indicates the loss weight, and This is used to balance numerical prediction accuracy with spatial phase consistency.

[0157] The real channel matrix and prediction channel matrix Vectorized into real channel vectors respectively and predicted channel vector The generalized cosine phase consistency loss Represented as:

[0158] ;

[0159] in, Represents the true channel vector The conjugate transpose of; Represents the magnitude of the complex inner product of the true channel vector and the predicted channel vector; Represents the true channel vector The second norm; Represents the predicted channel vector The second norm; This indicates that the expected value or average value of the sample is calculated.

[0160] More specifically, step S7 includes:

[0161] Keeping the network parameters of the continuous channel trajectory reconstruction network unchanged, the gradient of the joint optimization loss function with respect to the network parameters of the future channel state prediction network is calculated, and the network parameters of the future channel state prediction network are updated asynchronously based on the gradient; the updated network parameters of the future channel state prediction network are expressed as:

[0162] ;

[0163] in, This represents the network parameters of the future channel state prediction network before the asynchronous update begins; Indicates the delay after asynchronous processing. Complete the updated network parameters of the future channel state prediction network; Indicates the start time of the asynchronous update; This indicates the processing delay experienced by the asynchronous update from start to completion; The learning rate represents the network parameter update rate; This represents the joint optimization loss function calculated based on the network parameters of the network predicting the current future channel state. This represents the joint optimization loss function relative to the network parameters of the future channel state prediction network. The gradient;

[0164] During the asynchronous parameter update of the future channel state prediction network, the network parameters before the update are still used to perform future channel state prediction.

[0165] The future channel state prediction network with updated network parameters is used to predict the future channel state. Based on the new physical observations, the prediction error calculation, joint optimization loss function construction, and asynchronous parameter update of the second-level channel prediction network are performed, thus forming a closed-loop processing process in which physical channel observation, continuous channel trajectory reconstruction, future channel state prediction, downlink precoding, prediction error feedback, and network parameter update are interconnected.

[0166] Table 1 illustrates the reconstruction effect of the continuous channel trajectory reconstruction network on the uniformly sparse pilot observation data within the sliding time window in step S3. The maximum Doppler shift in Table 1 characterizes the time-varying nature of the channel caused by the movement of the mobile terminal. A larger maximum Doppler shift indicates a faster channel change. "Mixed" represents scenarios where the maximum Doppler shift changes over time. NMSE (Normalized Mean Square Error) indicates that a smaller value means the continuous historical channel trajectory is closer to the actual channel state. The table below illustrates this.

[0167] Table 1

[0168]

[0169] Under the conditions of CDL-A channel, signal-to-noise ratio of 20dB, and pilot spacing of 5ms, the present invention achieves low and stable NMSE under different maximum Doppler frequency shifts. This indicates that the sliding time window established in step S2 and the local multi-scale feature extraction, time-domain coding, and mask fusion in step S3 can reconstruct the channel state at unobserved time locations under high-speed time-varying channel conditions and retain physical observation values ​​at actual pilot observation locations, thereby providing continuous historical channel trajectories for future channel state prediction in step S4.

[0170] Table 2 illustrates the reconstruction performance of the continuous channel trajectory reconstruction network in step S3 under different signal-to-noise ratio (SNR) conditions when the maximum Doppler frequency shift is 90Hz. The horizontal columns represent different SNRs of the uniformly sparse pilot observation data, and the vertical column, NMSE, characterizes the normalized error between the reconstructed continuous historical channel trajectory and the actual channel state. See below:

[0171] Table 2

[0172]

[0173] As the signal-to-noise ratio (SNR) increases from −5dB to 30dB, the NMSE gradually decreases, indicating that after the pilot observation quality is improved, the continuous channel trajectory reconstruction network can more accurately recover the channel state at the unobserved time position within the sliding time window. At a lower SNR, the network can still utilize the timing information saved by the sliding time window in step S2, as well as the binary pilot position mask and mask fusion mechanism in step S3, to reduce the impact of noise on continuous historical channel trajectories, providing reliable input for the continuous prediction and periodic updates in steps S4 and S5.

[0174] Table 3 illustrates the adaptability of the continuous channel trajectory reconstruction network in step S3 to different 3GPP channel models under the conditions of a maximum Doppler frequency shift of 90Hz and a signal-to-noise ratio of 20dB; CDL-A, CDL-C, and CDL-D represent channel models with different multipath propagation characteristics, as follows:

[0175] Table 3

[0176]

[0177] Table 3 illustrates that this invention is not limited to a single propagation condition: the continuous channel trajectory reconstruction network continuously reconstructs uniform sparse pilot observation data under different channel models through input encoding, local multi-scale feature extraction, time-domain encoding, output projection, and mask fusion. The error is lower under the CDL-D condition, while the error is relatively larger under the CDL-C condition due to the greater complexity of multipath propagation; however, continuous historical channel trajectories usable in step S4 can be formed under all channel models.

[0178] Figure 2 The plot shows the normalized mean square error of different prediction schemes as the signal-to-noise ratio changes in a dynamic moving scenario; the horizontal axis represents the signal-to-noise ratio, and the vertical axis represents the normalized mean square error. The lower the value of the vertical axis, the closer the predicted channel state is to the actual physical channel state at the corresponding moment.

[0179] In the legend, "delay 5ms", "delay 10ms", and "delay 30ms" represent the asynchronous processing delays of 5ms, 10ms, and 30ms respectively when the future channel state prediction network is asynchronously updated according to step S7 of this invention. The asynchronous processing delay refers to the time taken to complete the background update of network parameters after obtaining the joint optimization loss function in step S6. Open-loop prediction represents a control scheme that does not use new physical channel observation data to periodically correct the prediction results; CT-Transformer, Neural ODE, Basic GRU, and Basic Transformer are all comparative prediction schemes.

[0180] The dashed box in the figure represents the locally magnified area under high signal-to-noise ratio, and the inset in the upper right corner shows the magnified result of this area. This figure illustrates that step S5 updates the sliding time window in the digital twin layer with new pilot observation data and uses the new physical observation data as the physical observation anchor point; step S7 asynchronously updates the future channel state prediction network without interrupting online inference; this can reduce the error accumulation caused by long-term open-loop prediction; the shorter the asynchronous processing delay, the better the updated network parameters match the current channel state.

[0181] Figure 3 The diagram illustrates the system capacity under different signal-to-noise ratios (SNRs) when performing downlink beamforming based on the predicted channel matrices obtained from various prediction schemes. The horizontal axis represents the SNR, and the vertical axis represents the system capacity, in bits per second (Hz). A higher value indicates a greater amount of information that can be transmitted per unit bandwidth. The upper "ideal capacity" axis represents the capacity reference value when beamforming is performed using the actual physical channel conditions at the corresponding time. This value reflects the theoretical upper limit and does not correspond to additional prediction method curves.

[0182] In the legend, "Outdated CSI" represents a control scheme that does not perform future channel state prediction but directly uses channel state information from an earlier time point; CT-Transformer, Neural ODE, Basic GRU, and Basic Transformer are control prediction schemes.

[0183] The inset in the upper left corner is a magnified view of a high signal-to-noise ratio (SNR) region, with the dashed box and connecting dashed lines indicating the magnified area. This figure corresponds to step S6: the base station performs singular value decomposition based on the predicted channel matrix and generates a downlink precoding matrix, which is then received and combined by the mobile terminal. The joint optimization of the loss function simultaneously constrains the accuracy of the channel numerical prediction and the spatial phase consistency of the complex channel, thereby reducing residual inter-stream crosstalk caused by prediction errors. Therefore, the method of this invention can achieve higher system capacity and more closely approximate the ideal capacity under medium-to-high SNR conditions.

[0184] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A digital twin channel prediction method for high mobile MIMO, characterized in that, The steps are as follows: S1: Construct a channel evolution model and determine the channel aging interval; obtain uniform sparse pilot observation data; S2: Construct a physical channel layer and a digital twin layer, transmit uniform sparse pilot observation data from the physical channel layer to the digital twin layer, and establish a sliding time window in the digital twin layer; S3: Use a continuous channel trajectory reconstruction network to reconstruct the continuous channel trajectory of uniform sparse pilot observation data within a sliding time window to obtain continuous historical channel trajectories. S4: Use a future channel state prediction network to predict the future channel state of continuous historical channel trajectories; S5: Periodically update the sliding time window in the digital twin layer, and cyclically perform continuous historical channel trajectory reconstruction, future channel state prediction and sliding time window update.

2. The digital twin channel prediction method for high mobile MIMO according to claim 1, characterized in that, Step S1 specifically includes: A channel evolution model is constructed to describe the correlation between channel states at adjacent time points in high-mobility multi-antenna communication scenarios. A first-order Gaussian-Markov relation is used to correlate the channel states at adjacent time points, expressed as: ； in, Indicates time The complex downlink channel matrix, and ; Indicates time The complex downlink channel matrix; Represents the set of complex numbers; Indicates the number of transmitting antennas at the base station; Indicates the number of receiving antennas in the mobile terminal; This represents the total time interval between the pilot observation time and the actual downlink transmission time; Indicates the relationship between the maximum Doppler frequency shift and the time interval. The relevant channel time correlation coefficient is used to characterize the degree of correlation between the channel states at two different times; Indicates time Relative to time Channel aging variation term; Determine the channel aging interval between the pilot observation time and the actual downlink transmission time, and express the total time interval corresponding to the channel aging interval as: ； in, The pilot probe interval represents a uniformly sparse pilot; This represents the total processing delay from the acquisition of pilot observation data to the execution of actual downlink transmission; The maximum Doppler frequency shift is determined based on the motion state of the mobile terminal, and combined with the total time interval. Determine the degree of channel time correlation; Acquire uniform sparse pilot observation data sent by the mobile terminal and received by the base station.

3. The digital twin channel prediction method for high mobile MIMO according to claim 1, characterized in that, Step S2 specifically includes: A physical channel layer is constructed, and uniform sparse pilot observation data is transmitted to the digital twin layer as the observation basis for the digital twin layer to characterize the current state of the physical channel layer. A digital twin layer is constructed, and a sliding time window corresponding to the current physical channel layer is established in the digital twin layer. The uniform sparse pilot observation data obtained in step S1 is stored in the sliding time window according to the pilot observation time.

4. The digital twin channel prediction method for high mobile MIMO according to claim 1, characterized in that, Step S3 specifically includes: A continuous channel trajectory reconstruction network is constructed. The continuous channel trajectory reconstruction network is the first-level channel reconstruction network in a two-level cascaded network. It is used to reconstruct the channel state of the unobserved time position within the sliding time window based on the uniform sparse pilot observation data and the corresponding pilot position information within the sliding time window, so as to obtain the continuous historical channel trajectory. The complex channel state information at each pilot observation time is separated into real and imaginary parts and arranged along the feature dimension to form a sparse pilot input tensor. Construct a binary pilot position mask based on the pilot positions corresponding to the uniform sparse pilot observation data. In locations where actual pilot observations exist, the corresponding mask element is set to 1; in locations where no actual pilot observations exist, the corresponding mask element is set to 0. Input the sparse pilot into the tensor With the binary pilot position mask The splicing is performed, and the splicing result is linearly projected to map the splicing result to the latent feature space; absolute position encoding is added to the features obtained by linear projection to preserve the chronological order of each time sampling position within the sliding time window; The features after adding absolute position encoding are subjected to concatenated one-dimensional residual dilated convolution processing to extract local multi-scale channel change features. The one-dimensional residual dilated convolution unit adopts different dilation rates to expand the temporal receptive field. Among them, the small dilation rate is used to extract fast fading and local phase change features, and the large dilation rate is used to extract long-term channel evolution features. Each one-dimensional residual dilated convolution unit sequentially performs one-dimensional dilated convolution, batch normalization, Gaussian error linear unit activation, and residual connection. The features processed by cascaded one-dimensional residual dilated convolution are input into a time-domain encoder based on a self-attention mechanism to establish long-range dependencies between different time positions within a sliding time window; then, output projection is performed to obtain the original continuous channel reconstruction result. ; The original continuous channel reconstruction result Perform mask fusion, retain the corresponding physical observation values ​​at the actual pilot observation locations, and use the reconstructed values ​​output by the first-level channel reconstruction network at the unobserved locations to obtain continuous historical channel trajectories, represented as: ； in, This represents the continuous historical channel trajectory obtained after mask fusion; Represents the binary pilot position mask Tensors with the same dimensions and all elements being 1; This represents the element-wise multiplication operation between elements at corresponding positions. This represents the inverse mask corresponding to the unobserved position, where the unobserved position is set to 1 and the actual pilot observation position is set to 0.

5. The digital twin channel prediction method for high mobile MIMO according to claim 4, characterized in that, Step S4 specifically includes: A future channel state prediction network is constructed. This network serves as the second-level channel prediction network in a two-stage cascaded network. Based on the continuous historical channel trajectories obtained in step S3, it extracts channel evolution features at different time scales and spatial correlation features between different antenna links. It then predicts the channel state at multiple future moments within a forward prediction interval corresponding to a pilot detection interval in a non-autoregressive manner. The time domain length of the forward prediction interval is the same as the pilot detection interval, and the prediction result includes the channel state corresponding to the next pilot observation moment. The continuous historical channel trajectories obtained in step S3 are arranged according to different antenna links, so that the historical channel state sequence of each antenna link within the sliding time window is input as an independent one-dimensional time series into the second-level channel prediction network. The historical channel state sequences of each antenna link are subjected to dual-path embedding processing of multi-scale convolution and linear mapping. The multi-scale convolution branch uses multiple one-dimensional convolution kernels of different scales to extract channel evolution features at different time scales. Small-scale one-dimensional convolution kernels are used to extract short-term rapid fading and phase change features, medium-scale one-dimensional convolution kernels are used to extract medium-scale change trends, and large-scale one-dimensional convolution kernels are used to extract slowly varying envelope features. The linear mapping branch performs linear mapping on the complete historical channel state sequence to preserve the overall channel evolution trend. By adaptively fusing the outputs of the two branches using learnable fusion parameters, a fused embedding feature is obtained. , is represented as: ； in, This represents the local multi-scale channel characteristics of the multi-scale convolution branch output; This represents the overall channel evolution characteristics of the linear mapping branch output; This represents the fused embedding feature obtained after adaptive fusion of the two branches; Represents the learnable fusion parameters; This represents a logistic activation function used to fuse the parameters. Mapped to fusion weights between 0 and 1; This represents the fusion weights corresponding to the linear mapping branches; The fused embedding features of each antenna link within the entire sliding time window are treated as a variable representation unit, and cross-variable self-attention processing is performed between different antenna links to utilize the spatial correlation between different antenna links in a co-propagation environment; for the ... The antenna link and the first For each antenna link, the attention weights are expressed as follows: ； The value vectors of different antenna links are weighted and aggregated according to the attention weights to obtain the first... The spatial compensation characteristics of an antenna link are expressed as follows: ； in, Indicates the first The antenna link pair of the first Attention weights for each antenna link; Indicates the first The query vector for each antenna link; Indicates the first The key vector of each antenna link; Indicates the first The value vector of each antenna link; In the process of summation, the first... The key vector of each antenna link; Indicates the transpose operation; This represents the feature dimension of the key vector; This indicates the number of antenna links participating in the intervariate self-attention process; Indicates the summation index; Represents an exponential function; Indicates the first Spatial compensation features obtained by transvariable self-attention processing of individual antenna links; The features processed by intervariate self-attention are output-mapped, and the final residuals are connected to generate the future channel state prediction results, as follows: ； in, This indicates the channel state at the end of the sliding time window of the continuous historical channel trajectory; Indicates the length of the historical channel state sequence within the sliding time window; This represents the future channel change predicted by the second-level channel prediction network based on historical channel evolution patterns. This indicates the prediction result of the future channel state.

6. The digital twin channel prediction method for high mobile MIMO according to claim 5, characterized in that, Step S5 specifically includes: The sliding time window is moved forward by one pilot detection interval along the time direction. The earliest historical pilot observation data in the sliding time window is removed. The new pilot observation data and its corresponding pilot position information are stored in the sliding time window, and the binary pilot position mask is updated simultaneously, so that the updated sliding time window continues to maintain a uniform pilot sampling structure. Based on the updated sliding time window, step S3 is re-executed, and the continuous channel trajectory is reconstructed on the updated pilot observation data using the continuous channel trajectory reconstruction network to obtain the updated continuous historical channel trajectory; then, based on the updated continuous historical channel trajectory, step S4 is re-executed, and the updated future channel state prediction result is obtained using the future channel state prediction network. During each sliding time window update, the new pilot observation data is used as the new physical observation anchor point, so that the continuous historical channel trajectory and the prediction results of future channel state are periodically constrained by the latest physical channel observation. This reduces the accumulation of prediction errors caused by long-term open-loop extrapolation and keeps the continuous channel reconstruction and future channel state prediction process within a wide stable range corresponding to the current channel statistical characteristics.

7. The digital twin channel prediction method for high mobile MIMO according to claim 6, characterized in that, The method further includes: S6: Determine the prediction error based on the future channel state prediction results and new pilot observation data, and construct a joint optimization loss function that simultaneously constrains the channel numerical prediction accuracy and spatial phase consistency based on the prediction error; S7: Based on the joint optimization loss function obtained in step S6, keep the network parameters of the continuous channel trajectory reconstruction network unchanged, and asynchronously update the network parameters of the future channel state prediction network; during the parameter update, continue to use the network parameters before the update for future channel state prediction, and after the parameter update is completed, use the updated network parameters for subsequent future channel state prediction.

8. The digital twin channel prediction method for high mobile MIMO according to claim 7, characterized in that, Step S6 specifically includes: The prediction channel matrix corresponding to the future channel state prediction results obtained in step S4 Perform singular value decomposition, represented as: ； in, This represents the unitary matrix composed of the left singular vectors of the prediction channel matrix; This represents a matrix composed of the singular values ​​of the prediction channel matrix; This represents a unitary matrix composed of the right singular vectors of the prediction channel matrix; the base station uses this matrix... Generate a downlink precoding matrix, which the mobile terminal then utilizes. Perform receive merging to obtain the receive merge matrix; When new pilot observation data acquired periodically in step S5 arrives, and its pilot observation time corresponds to the corresponding prediction time in the previous future channel state prediction results, the physical channel observation reference state at the corresponding time is obtained based on the new pilot observation data, and the prediction error matrix between the observation reference channel matrix and the prediction channel matrix corresponding to the physical channel observation reference state is calculated, as follows: ； in, Represents the prediction error matrix; This represents the observation reference channel matrix corresponding to the time of the new pilot observation data; This represents the prediction channel matrix at the corresponding time point; Applying the precoding matrix and receive-combining matrix obtained based on the predicted channel matrix to the real physical channel yields the actual equivalent baseband signal, expressed as: ； in, This represents the equivalent baseband received signal vector after merging; Indicates the downlink transmitted symbol vector; Represents the noise vector of the mobile terminal; This represents the desired signal component formed after spatial flow decoupling based on the predicted channel matrix; Represents the prediction error matrix The resulting residual inter-flow crosstalk components; A joint optimization loss function is constructed based on the prediction error, which simultaneously constrains the channel numerical prediction error and the phase direction deviation of the complex channel in vector space, and is expressed as: ； in, Represents the joint optimization loss function; This represents the mean squared error loss, used to measure the numerical difference between the actual channel state and the predicted channel state. It represents the generalized cosine phase consistency loss, which is used to measure the directional difference between the actual channel state and the predicted channel state in the complex vector space; Indicates the loss weight, and This is used to balance numerical prediction accuracy with spatial phase consistency. The real channel matrix and prediction channel matrix Vectorized into real channel vectors respectively and predicted channel vector The generalized cosine phase consistency loss Represented as: ； in, Represents the true channel vector The conjugate transpose of; Represents the magnitude of the complex inner product of the true channel vector and the predicted channel vector; Represents the true channel vector The second norm; Represents the predicted channel vector The second norm; This indicates that the expected value or average value of the sample is calculated.

9. The digital twin channel prediction method for high mobile MIMO according to claim 8, characterized in that, Step S7 specifically includes: Keeping the network parameters of the continuous channel trajectory reconstruction network unchanged, the gradient of the joint optimization loss function with respect to the network parameters of the future channel state prediction network is calculated, and the network parameters of the future channel state prediction network are updated asynchronously based on the gradient; the updated network parameters of the future channel state prediction network are expressed as: ； in, This represents the network parameters of the future channel state prediction network before the asynchronous update begins; Indicates the delay after asynchronous processing. Complete the updated network parameters of the future channel state prediction network; Indicates the start time of the asynchronous update; This indicates the processing delay experienced by the asynchronous update from start to completion; The learning rate represents the network parameter update rate; This represents the joint optimization loss function calculated based on the network parameters of the network predicting the current future channel state. This represents the joint optimization loss function relative to the network parameters of the future channel state prediction network. The gradient; During the asynchronous parameter update of the future channel state prediction network, the network parameters before the update are still used to perform future channel state prediction. The future channel state prediction network with updated network parameters is used to predict the subsequent future channel state.

Images