A physical conditioning channel state information autoregressive prediction method based on a large language model

Abstract

The application discloses a physical conditioning channel state information autoregressive prediction method based on a large language model, comprising: performing continuous characterization processing on original channel state information matrices collected by a wireless channel; extracting physical metadata; inputting a physical conditioning input sequence into a pre-trained large language model to complete autoregressive modeling, and outputting a predicted channel state information matrix; introducing a difficulty perception course training strategy, dividing difficulty levels of training samples according to the physical metadata, and adopting a phased dynamic sampling mode to optimize the pre-training large language model training process; and using the trained pre-training large language model to combine historical channel state information data and physical metadata of an actual scene to complete autoregressive prediction of channel state information at a future time. The application effectively solves the characterization bottleneck caused by implicit physical information in traditional CSI prediction technology, greatly reduces the modeling burden of the model, and improves the basic accuracy of CSI prediction.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to an autoregressive prediction method for physically conditional channel state information based on a large language model. Background Technology

[0002] Accurate prediction of Channel State Information (CSI) is one of the core technologies of modern wireless communication systems. It is the foundation for key functions such as adaptive beamforming, resource scheduling, and link adaptation in 5G-Advanced and 6G networks, directly determining the transmission efficiency, resource utilization, and dynamic adaptability of wireless communication systems. With the diversification of wireless communication system deployment environments, the heterogeneity of carrier frequencies, propagation scenarios, and user mobility has significantly increased. How to ensure that the CSI prediction model maintains stable performance within a shared frequency band while achieving effective out-of-distribution generalization with zero samples across frequencies has become an urgent problem to be solved in this field.

[0003] Existing CSI prediction solutions can be broadly categorized into three types, but all suffer from insurmountable limitations, failing to meet the robustness and generalization requirements under heterogeneous propagation conditions. First, traditional model-driven prediction methods construct parameterized models based on the assumption of stable channel statistical characteristics, requiring fine-tuning of model parameters. In scenarios with drastic channel dynamics or changes in the deployment environment, the models are prone to failure, leading to a significant drop in prediction accuracy. Second, while data-driven prediction methods improve average prediction accuracy by relying on large amounts of data, they exhibit significant configuration specificity. When carrier frequency, propagation scenario, or antenna configuration changes, the model needs retraining, resulting in poor generalization ability and high annotation and training costs, leading to insufficient adaptability. Finally, the emerging CSI prediction methods and wireless fundamental models based on Large Language Models (LLM) have improved cross-scenario generalization performance to some extent, but their underlying mechanism for maintaining robustness under frequency changes remains unclear. The prediction performance of these models under short context windows and worst-case carrier bands remains unstable, making it difficult to adapt to complex real-world deployment scenarios.

[0004] A deeper analysis reveals that the performance limitations of existing CSI prediction methods stem from two coupled issues: the representation bottleneck caused by the lack of explicit integration of physical metadata into the input, and insufficient learning of complex physical scenarios due to non-targeted training sampling. This makes it impossible for existing models to achieve stable predictions within shared frequency bands, and also hinders zero-sample cross-frequency generalization, making it difficult to meet the actual deployment requirements of 5G-Advanced and 6G networks. Therefore, an innovative CSI prediction method is urgently needed to address these technical deficiencies from both the input and training layers without altering the model's core architecture, thereby improving the model's robustness and generalization capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an autoregressive prediction method for physically conditional channel state information based on a large language model, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A physical conditional channel state information autoregressive prediction method based on a large language model includes: S1. Perform continuous representation processing on the original channel state information matrix collected by the wireless channel, and convert the complex-valued channel state information data into continuous channel state information embedding features that are adapted to the pre-trained large language model sequence modeling. S2. Extract physical metadata, convert it into a prompt token through the physical conditional prompt coding module, and concatenate it to the front end of the sequence of the continuous channel state information embedding features to form a complete physical conditional input sequence. S3. The physical conditional input sequence is fed into a pre-trained large language model to complete autoregressive modeling and output the predicted channel state information matrix. The pre-trained large language model adopts an end-to-end joint training method, uses a combined loss function consisting of mean square error loss and complex angle consistency loss, masks the loss calculation of physical cue tokens, and only optimizes the loss for channel state information sequence features. S4. Introduce a difficulty-aware course training strategy, divide training samples into difficulty levels according to physical metadata, and optimize the pre-trained large language model training process by adopting a phased dynamic sampling method. S5. Using the pre-trained large language model that has been trained, combined with historical channel state information data and physical metadata from real-world scenarios, autoregressive prediction of channel state information at future moments is completed.

[0007] Furthermore, the continuous representation processing in S1 specifically involves: decomposing the complex-valued channel state information matrix into real and imaginary parts and then concatenating them into vectors; performing zero-mean-unit-variance normalization on the concatenated data to obtain a real-valued feature vector of uniform scale; projecting this real-valued feature vector onto the hidden dimension space of the pre-trained large language model through a lightweight linear mapping layer to generate continuous channel state information embedding features; and having the learnable parameters of the lightweight linear mapping layer participate in end-to-end joint training.

[0008] Furthermore, the physical metadata in S2 includes three core parameters: carrier frequency, propagation scenario type, and user movement speed. The working process of the physical conditional cue coding module is as follows: the carrier frequency and user movement speed are normalized and then fused into Doppler proxy features, and Doppler cue tokens are generated through multi-scale Fourier feature encoding; the propagation scenario type is input into the learnable scenario embedding table and converted into a scenario category cue token; the Doppler cue token and the scenario category cue token are concatenated in a fixed order to form a physical conditional prefix, and then concatenated with the sequence of continuous channel state information embedding features to form a physical conditional input sequence.

[0009] Furthermore, the Doppler proxy feature is obtained by multiplying the carrier frequency normalization result with the user movement speed normalization result. The frequency coefficient scale of the multi-scale Fourier feature encoding includes {2,5,10,25,50}, and the number of generated Doppler cue tokens is 5.

[0010] Furthermore, the autoregressive modeling in S3 specifically involves: using a general pre-trained large language model backbone, relying on its long sequence modeling capability to generate channel state information embedding features for multiple future time steps through point-by-point autoregression; mapping the channel state information embedding features back to the physical domain through a lightweight linear regression head, and restoring them to a prediction matrix consistent with the original channel state information format and scale after dimensional reshaping and inverse normalization processing.

[0011] Furthermore, the combined loss function L in S3 pred Satisfy: L pred =λ1L MSE +λ2L angle Where λ1 and λ2 are weighting coefficients, and λ1 + λ2 = 1; L MSE L represents the complex mean square error, used to constrain the accuracy of CSI amplitude reconstruction. angle This is the phase angle error, used to ensure that the predicted CSI is aligned with the complex phase of the true CSI.

[0012] Furthermore, the difficulty-aware course training strategy in S4 is as follows: a comprehensive difficulty score of the samples is calculated based on the scarcity of carrier frequencies, user mobility, and scene propagation complexity, and the training samples are divided into three levels: easy, medium, and difficult; dynamic sampling is carried out in three stages: in the early stage of training, easy samples are the main focus, with a small number of medium and difficult samples added; in the middle stage of training, the sampling ratio of medium and difficult samples is gradually increased; in the later stage of training, a near-uniform sampling method is adopted with a slight bias towards difficult samples; and in the early stage of training, an auxiliary scheduling strategy that prioritizes sampling samples with long context windows is adopted simultaneously.

[0013] Furthermore, the scores for each dimension of the overall difficulty score of the sample are normalized to the [0,1] interval.

[0014] Furthermore, the autoregressive prediction process of S5 is as follows: collect historical channel state information data and corresponding physical metadata of the actual scenario, complete the construction of continuous representation and physical conditional input sequence according to the steps of S1 and S2; generate future multi-step channel state information embedding features from the backbone of the general pre-trained large language model through autoregression, and after linear regression head mapping, dimension reshaping and inverse normalization, output the channel state information prediction result that meets the requirements of wireless communication system.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1) This invention uses a physical conditional prompting coding module to explicitly inject key physical metadata such as carrier frequency, scene type, and user mobility into the model input, effectively solving the representation bottleneck caused by the implicit physical information in traditional CSI prediction technology. It successfully separates the two major tasks of physical scene inference and channel timing prediction, allowing the backbone of the large language model to focus on channel evolution modeling under explicit physical conditions, greatly reducing the modeling burden of the model and improving the basic accuracy of CSI prediction.

[0016] 2) This invention introduces a difficulty-aware course training strategy, which performs difficulty stratification and phased dynamic sampling of training samples based on physical metadata. It focuses on strengthening the learning weights of complex physical scenarios such as rare frequencies, high mobility, and non-line-of-sight, effectively solving the problem of insufficient exposure of difficult scenarios in the traditional pure random sampling method, and significantly improving the prediction stability of the model in the shared frequency band and the adaptability of the worst carrier frequency band.

[0017] 3) This invention adopts a fully continuous domain CSI processing method throughout the process, without using discretized tokens, and fully preserves the original physical characteristics and detailed information of the wireless channel state information. At the same time, through a combined loss function composed of mean square error loss and complex angle consistency loss, it takes into account the dual prediction accuracy of CSI amplitude and phase, effectively avoids phase distortion, and improves the availability of prediction results in actual communication systems.

[0018] 4) The core innovation of this invention focuses on the construction of the model's input layer and the selection of training strategies. It does not modify or reconstruct the backbone architecture of the pre-trained large language model, thus avoiding additional computational overhead and computing power costs. The technical solution is easy to deploy in wireless communication systems.

[0019] 5) This invention enables the model to accurately capture channel evolution characteristics that are highly aligned with the physical laws of wireless propagation through the collaborative optimization of physical conditional input and difficulty-aware training, significantly improving the physical interpretability of the model. At the same time, the model can capture cross-frequency universal channel patterns under the guidance of explicit physical metadata, effectively breaking through the generalization barrier caused by frequency heterogeneity, achieving reliable zero-sample cross-frequency distribution out-generalization, and meeting the CSI prediction requirements in heterogeneous communication scenarios. Attached Figure Description

[0020] Figure 1 The overall framework diagram of the physical conditional CSI autoregressive prediction method based on a large language model provided in the embodiments of the present invention is shown.

[0021] Figure 2 This is a schematic diagram of the CSI matrix continuous representation processing flow provided in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the details of physical metadata encoding and prompt token concatenation provided in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of sample division and dynamic sampling for difficulty-aware course training provided in an embodiment of the present invention.

[0024] Figure 5 The following is a schematic diagram for verifying the consistency between the attention mechanism provided in this embodiment of the invention and Clarke-style channel coherence theory: (a) is a heat map of the average effective backtracking length of attention on a two-dimensional plane composed of user movement speed and carrier frequency; (b) is the cumulative probability distribution curve of median lag of attention under different Doppler levels; (c) is the curve of the average normalized mean square error changing with the context length under historical context truncation scanning. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention proposes a physical conditional channel state information (CSI) autoregressive prediction method based on a large language model, applicable to 5G-Advanced and 6G wireless communication systems. Its design goal is to improve the robustness and generalization ability of CSI prediction under heterogeneous frequency, scenario, and mobility conditions through the collaborative optimization of physical conditional coding at the input layer and difficulty-aware scheduling at the training layer. Simultaneously, it aligns the model's attention mechanism with the physical laws of wireless channels, enabling the model to achieve stable prediction within shared frequency band distributions and effective generalization outside zero-sample frequency shift distributions while maintaining the backbone architecture. Furthermore, using this method for end-to-end model training allows for direct integration into downstream modules of existing wireless communication systems, such as adaptive beamforming and resource scheduling, realizing the engineering implementation of CSI prediction. The specific components and working principles of this scheme will be detailed below.

[0027] like Figure 1As shown, the present invention provides an autoregressive prediction method for physically conditional channel state information based on a large language model, comprising: S1. Perform continuous representation processing on the raw channel state information matrix acquired from the wireless channel, converting the complex-valued channel state information data into continuous embedding features adapted for sequence modeling of a pre-trained large language model. This step specifically includes: Continuous representation processing of the original channel state information matrix: The core lies in achieving modal adaptation between the complex-valued channel state information physical signal and the embedding space of the pre-trained large language model. Through a fully continuous domain feature processing method, discrete tokenization operations are avoided throughout the process, fully preserving the original physical characteristics and detailed information of the channel state information data. This effectively avoids the impact of quantization errors and codebook collapse on the accuracy of subsequent model building. The specific processing flow and the connection between each stage are as follows: Figure 2 As shown, Figure 2 The paper intuitively demonstrates the complete transformation process from the original complex-valued channel state information matrix to the final continuous embedded features. First, the original channel state information matrix of the wireless communication channel at time step t is collected. This matrix is ​​a complex-valued matrix, which can accurately characterize the channel transmission characteristics between the transmitting and receiving antennas at the current moment. This complex-valued channel state information matrix is ​​decomposed into a real part matrix and an imaginary part matrix. Vectorization operations are performed on the real and imaginary parts matrices respectively, converting the two-dimensional matrix data into one-dimensional vectors. Then, the resulting one-dimensional real and imaginary vectors are concatenated column-wise to form the original form of the real-valued feature vector. This operation completes the transformation from complex-valued data to real-valued data, eliminating the fundamental modal differences between complex-valued modes and the real-valued embedding space of large language models.

[0028] Obtaining continuous channel state information embedding features: The concatenated original real-valued feature vectors are normalized to zero mean and unit variance to eliminate numerical scale bias caused by differences in physical environment of channel state information data under different carrier frequencies, propagation scenarios, and antenna configurations. This allows channel state information data in heterogeneous scenarios to form a feature representation of a uniform scale, thus adapting to the sequence modeling logic of the large language model and ensuring that the model can learn the temporal evolution law of channel state information data in different scenarios in a unified manner. After normalization, a lightweight linear mapping layer consisting of learnable weight matrices and bias terms is set. The lightweight linear mapping layer is a neural network layer that performs linear transformation operations, projecting the normalized uniform-scale real-valued feature vectors onto the hidden dimension space of the pre-trained large language model to generate continuous channel state information embedding features. All learnable parameters of this linear mapping layer will participate in end-to-end joint training with the parameters of the subsequent large language model backbone network and other modules. During training, it will be continuously optimized and updated to ensure that the expressive power of the generated CSI continuous embedding features is adapted to the feature processing logic of the large language model backbone, realizing the effective conversion of CSI physical features to model-recognizable features.

[0029] This step employs a fully continuous domain processing strategy throughout, without performing any discretization or quantization operations on the CSI data. It fully preserves the core physical information contained in the CSI data, such as amplitude and phase, which characterize the transmission properties of the wireless channel, providing a high-quality feature foundation for the autoregressive prediction of the subsequent model.

[0030] S2. Extract physical metadata and convert it into cue tokens using the physical conditional cue coding module. These tokens are then concatenated to the front end of the continuous channel state information (CSI) embedding feature to form a complete physical conditional input sequence. In other words, physical metadata is converted into cue tokens and placed at the beginning of the CSI embedding feature to construct a complete input sequence with physical prior information, which is then fed to the model. This step specifically includes: Physical metadata extraction and feature encoding: The core lies in encoding directly collectable physical environment parameters in wireless communication scenarios into cue tokens adapted to large language models, injecting explicit physical scene priors into the model, and solving the bottleneck of implicit physical information representation in traditional CSI prediction. The specific encoding process and feature transformation details are as follows: Figure 3 As shown, Figure 3The document visually demonstrates the complete generation process from physical metadata to various prompt tokens. First, from actual wireless communication deployment scenarios, three core physical metadata types strongly correlated with channel dynamics are extracted simultaneously: carrier frequency (fc), propagation scenario type (s), and user movement speed (v). These three metadata types are deployment and environmental parameters that can be obtained in real time during the operation of the wireless communication system, accurately characterizing the physical scenario features of the channel and serving as the core basis for constructing physical conditional prompts. Based on Clarke-style channel coherence theory, the Doppler spread of the channel is positively correlated with the carrier frequency and user movement speed, which are key physical factors determining the channel's temporal evolution characteristics. Therefore, the carrier frequency (fc) and user movement speed (v) are first normalized separately. A truncation function is used to strictly constrain the normalization results to the [0,1] interval, effectively avoiding interference from extreme value data on feature fusion. Then, the normalization results are multiplied and fused into a single-scalar Doppler proxy feature, fully preserving the monotonic dependence of carrier frequency and user movement speed on channel dynamics. The Doppler surrogate feature, derived from channel observation information, represents a substitute feature for the channel's Doppler dynamic characteristics. This Doppler surrogate feature is then input into a multi-scale Fourier feature encoding module. Multi-scale frequency encoding transforms the single scalar feature into a high-dimensional frequency feature capable of representing the channel dynamics in different Doppler domains. A lightweight projection layer then maps this feature to generate a Doppler cue token, which is an encoding of the Doppler surrogate feature and serves as a cue token representing the channel's Doppler dynamic attributes. Simultaneously, a learnable scene embedding table matching the hidden dimension space of the pre-trained large language model is constructed. Discrete propagation scene types *s* are input into this embedding table as indices. Scene category cue tokens are obtained through table lookup operations. This scene embedding table undergoes end-to-end updates and optimizations throughout model training, enabling adaptive learning of features from different propagation scenes. The parameters of the learnable scene embedding table are updated iteratively with network training and serve as a lookup table for embedding vectors corresponding to various wireless scenes. The scene category cue token is an encoding of the scene embedding vector and carries prior information about the wireless scene category.

[0031] The frequency coefficient scale of the multi-scale Fourier feature encoding includes {2,5,10,25,50}, and the number of generated Doppler cue tokens is 5.

[0032] Physically conditional input sequence construction: The generated Doppler cue tokens and scene category cue tokens are concatenated in a fixed logical order to form a structured physical conditional prefix. This physical conditional prefix, composed of at least one type of physical cue token, is concatenated to the preceding token subsequence of the channel embedding feature. This prefix is ​​then directly concatenated to the beginning of the continuous channel state information embedding feature sequence obtained in S1, constructing a complete physical conditional input sequence. This concatenation method ensures the pre-constraint effect of physical scene information in the input sequence and fully matches the sequence input format specification of the pre-trained large language model, allowing it to be directly fed into the model for subsequent autoregressive modeling. Figure 3 The code clearly illustrates the concatenation logic and data flow path of the physical conditional prefix and the CSI continuous embedding feature sequence, intuitively demonstrating the construction process of the complete input sequence. The physical conditional input sequence constructed in this step deeply integrates explicit physical scenario constraints with CSI temporal features, achieving task separation between physical scenario inference and channel temporal prediction. This allows the backbone of the large language model to focus on modeling the channel evolution law under explicit physical conditions, significantly reducing the modeling burden. At the same time, it provides key physical feature support for the model to achieve cross-frequency and cross-scenario generalized prediction.

[0033] S3. The physically conditionalized input sequence is fed into a pre-trained large language model to complete autoregressive modeling, outputting a predicted channel state information matrix. The pre-trained large language model employs an end-to-end joint training method, using a combined loss function consisting of mean square error loss and complex angle consistency loss. The loss calculation for the physical cue token is masked, and loss optimization is performed only on the channel state information sequence features. The complex angle consistency loss is used as a regularization term to constrain the deviation between the reconstructed complex angles of the channel state information and the true complex angles. The physical cue token is obtained by encoding and transforming physical metadata and is used to carry the cue token for carrying the prior physical information of the wireless channel.

[0034] This step specifically includes: Pre-trained large language model autoregressive feature modeling: The core lies in leveraging the mature long-sequence temporal modeling capabilities of pre-trained large language models to learn the temporal evolution patterns of historical CSI continuous embedding features, thereby achieving autoregressive generation of future CSI embedding features. In this embodiment, Qwen3-4B-Base is selected as the backbone of the pre-trained large language model. The complete physically conditionalized input sequence constructed by S2 is directly fed into this backbone network. This model is built on the Transformer architecture and has a powerful ability to capture long contextual features. It can effectively mine the temporal correlation information of CSI continuous embedding features in the physically conditionalized input sequence. Using a pointwise autoregressive generation method, it sequentially predicts the channel state information embedding features of multiple future time steps. In this embodiment, the prediction step size K=12 is set to meet the actual needs of dynamic adaptation in wireless communication, achieving accurate inference from historical CSI temporal features to future CSI temporal features. After the model outputs future CSI embedding features, a lightweight linear regression head matching the hidden dimensions of the large language model is set to map the CSI embedding features output by the model back to the physical domain, obtaining real-valued prediction features. Then, a dimension reshaping operation is performed on the real-valued prediction features to restore the one-dimensional features to a matrix form that is completely consistent with the dimensions and format of the original CSI matrix. Then, an inverse normalization process is performed. This inverse normalization process is the inverse operation of the normalization process of the original CSI data in S1. It can restore the predicted features output by the model to a result that is consistent with the numerical scale of the real CSI data. Finally, a CSI prediction matrix that is completely consistent with the format, scale, and physical meaning of the original CSI matrix is ​​obtained.

[0035] End-to-end joint training and loss optimization: All learnable parameters of the lightweight linear mapping layer of S1, the physical conditional cue encoding module of S2, the Qwen3-4B-Base pre-trained large language model backbone network, and the linear regression head of this step are included in the training scope, and an end-to-end joint training method is used for overall optimization. During training, a combined loss function consisting of a weighted average squared error loss and a complex angle consistency loss is designed, with the core calculation formula being L. pred =λ1L MSE +λ2L angle Where λ1 and λ2 are weighting coefficients, λ1 + λ2 = 1, and in this embodiment, the angle loss weight is set to λ2 = 0.5. MSE L represents the complex mean square error, used to constrain the accuracy of CSI amplitude reconstruction. angleThe phase angle error is used to ensure the complex phase alignment between the predicted CSI and the true CSI. Mean squared error loss is used to strictly constrain the reconstruction accuracy of the CSI data amplitude, while complex angle consistency loss ensures accurate alignment of the predicted CSI and the true CSI in complex phase, avoiding phase distortion caused by training with a single loss function. This ensures that the model's prediction results simultaneously meet the physical accuracy requirements of amplitude and phase. A loss masking strategy is also employed: during training, only the prediction results corresponding to the continuously embedded feature sequences of CSI are calculated for loss and backpropagation is performed. The physical cue token part does not participate in any loss calculation or parameter update, preventing the model from relying on physical metadata to form a simple optimization path. This ensures that the model truly learns the temporal evolution of the wireless channel, rather than simply fitting the superficial correlation between physical metadata and CSI data. Model training uses the AdamW optimizer to perform parameter updates, and an early stopping training strategy is implemented. Training is immediately stopped when the normalized mean squared error (NMSE) on the validation set does not decrease for several consecutive rounds, ensuring both the training effect and effectively avoiding overfitting.

[0036] S4. Introduce a difficulty-aware course training strategy. Difficulty levels are assigned to training samples based on physical metadata, and a phased dynamic sampling method is used to optimize the pre-trained large language model training process. This step specifically includes: Training sample difficulty quantification and stratification: The core lies in the fine-grained difficulty classification of training samples based on physical metadata, solving the problem of insufficient exposure of complex physical scene samples in traditional pure random sampling training methods. Its specific sample classification logic and hierarchical basis are as follows: Figure 4 As shown, Figure 4 The document visually demonstrates the complete process from sample difficulty scoring to difficulty level classification. First, based on physical metadata, a comprehensive difficulty score is calculated for each sample in the training set. Carrier frequency rarity, user mobility, and scene propagation complexity are selected as core scoring dimensions. Additional dimensions such as context window length and signal-to-noise ratio / noise level can be incorporated according to actual needs. The scores for each dimension are normalized, strictly limiting the scores to the [0,1] interval. Then, reasonable weighting coefficients are set for each dimension, and a weighted sum is obtained to obtain the comprehensive difficulty score for each training sample, ensuring that the difficulty score accurately represents the learning difficulty of the physical scene corresponding to the sample. Subsequently, a quantile binning strategy is adopted, sorting all samples in the training set from low to high based on their comprehensive difficulty scores. The 33rd and 66th percentiles are selected as difficulty classification thresholds, precisely dividing all training samples into three difficulty levels: easy, medium, and hard. This achieves refined difficulty stratification of the training samples, laying the foundation for subsequent phased dynamic sampling.

[0037] Among them, carrier frequency scarcity is used to characterize the learning difficulty of a certain carrier frequency point during training and the degree of insufficient sample exposure. The carrier frequency scarcity quantification method is to normalize the sample carrier frequencies within the training set carrier library using the normalization method. f =(fc-f min ) / (f max -f min ), D f f represents the rarity of carrier frequencies. min with f max The minimum and maximum values ​​are taken from the carrier set (carrier library) actually used in the training set. The value increases with frequency, representing both the physical challenges caused by Doppler enhancement and the insufficient exposure due to the low proportion of high-frequency samples.

[0038] The scene propagation complexity is calculated by weighting four normalized descriptors—direct path state, latency spread, shadow fading standard deviation, and multipath cluster number—to obtain the scene difficulty Ds. Scenes with non-line-of-sight and rich multipath elements score higher.

[0039] The comprehensive difficulty score is calculated by weighting and summing five normalized sub-items: frequency, movement speed, scene, context length, and noise, to obtain the difficulty score D for each sample. Then, all samples are divided into three levels of sample buckets: easy, medium, and difficult, according to the quantile threshold.

[0040] Phased dynamic sampling and training optimization: The overall training process of the model is divided into three stages. A dynamic mixed sampling method is used to adjust the sampling ratio of sample buckets of different difficulty levels, so as to achieve progressive learning of the model from easy to difficult. At the same time, an auxiliary context scheduling strategy is used. Figure 4The document clearly marks the sampling ratios and adjustment logic for each training stage. In the initial stage of training, which is the basic convergence phase, samples from simple sample buckets are primarily used, with appropriate additions of samples from medium and difficult sample buckets to ensure basic scene coverage. Simultaneously, samples with a context window length L≥12 are prioritized, accounting for no less than 50%, stabilizing the early training process. In the middle stage of training, which is the transition learning phase, the sampling ratio of simple sample buckets is gradually reduced linearly while the sampling ratio of medium and difficult sample buckets is increased, smoothly transitioning to a full data distribution sampling method. This allows the model to gradually adapt to learning complex physical scenes, while gradually restoring random sampling with the full context length to achieve a steady improvement in prediction accuracy. In the later stage of training, which is the reinforcement and optimization phase, a near-uniform sampling method is used to cover all difficulty sample buckets, while slightly favoring the sampling of difficult sample buckets. This ensures the model learns the entire scene while strengthening the learning of hard physical domain samples such as rare frequencies, high mobility, and non-line-of-sight samples, while avoiding the distribution drift problem in the later stages of model training caused by excessive favoritism towards difficult samples. By employing this difficulty-aware course training strategy, the training sampling distribution of the model is rebalanced, significantly improving the model's ability to learn from complex physical scenarios. This effectively enhances the model's prediction robustness under heterogeneous frequency, scenario, and mobility conditions. Moreover, this strategy only applies to the model training phase and requires no additional operations during inference, thus not increasing the computational overhead of model deployment.

[0041] S5. Using the pre-trained large language model, combined with historical channel state information data and physical metadata from real-world scenarios, perform autoregressive prediction of channel state information for future timeframes. This step specifically includes: Real-world data acquisition and input sequence construction: The core lies in deploying the trained model to real-world wireless communication scenarios, achieving real-time autoregressive prediction of CSI. The entire inference process maintains the core architecture of the Qwen3-4B-Base pre-trained large language model without any additional computational overhead or model reconstruction operations, and can be directly integrated into existing 5G-Advanced and 6G wireless communication systems. First, historical CSI data of length L is collected in real-time from the actual wireless communication scenario to be predicted, according to the system's channel detection frequency. In this embodiment, L can be flexibly selected from different lengths such as 4, 8, 12, 16, and 20 to adapt to the observation data conditions of different scenarios. Simultaneously, the physical metadata corresponding to this historical CSI data is collected, including carrier frequency, propagation scenario type, and user movement speed, strictly ensuring a one-to-one correspondence between the physical metadata and the historical CSI data in terms of time and scenario. The collected historical CSI data undergoes complete continuous representation processing according to steps S1, sequentially performing complex-valued matrix decomposition and concatenation, zero-mean-unit-variance normalization, and lightweight linear mapping to generate a CSI continuous embedding feature sequence consistent with the training phase format. Simultaneously, the synchronously collected physical metadata undergoes complete physical conditional cue encoding according to steps S2, sequentially performing Doppler surrogate feature calculation, multi-scale Fourier feature encoding to generate Doppler cue tokens, and scene embedding lookup table to generate scene category cue tokens. The two types of cue tokens are then concatenated into a physical conditional prefix, which is appended to the front end of the CSI continuous embedding feature sequence to form a physical conditional input sequence that can be directly fed into the model.

[0042] CSI Autoregressive Prediction and Results Output: The constructed physical conditional input sequence is fed into the trained model. Under explicit physical constraints, the model generates CSI embedding features for multiple future time steps in an autoregressive manner, following the channel temporal evolution patterns learned during training. The model's output future CSI embedding features are then fed into a linear regression head, mapped back to the physical domain to obtain real-valued prediction features. After dimensionality reshaping and inverse normalization, a CSI prediction matrix for multiple future time steps is obtained. This prediction matrix is ​​completely consistent with the actual collected CSI data in format, scale, and physical meaning, and can be directly recognized and used by the wireless communication system. Finally, the generated CSI prediction matrix is ​​output in real-time to downstream core functional modules such as adaptive beamforming, resource scheduling, and link adaptation, according to the interface specifications of existing wireless communication systems. This allows the wireless communication system to dynamically and adaptively allocate channel resources based on accurate CSI prediction results, effectively improving the system's transmission efficiency, resource utilization, and dynamic adaptation capabilities.

[0043] To verify the physical interpretability of the model in this invention, avoid the black-box defects caused by purely data-driven modeling, and ensure that the model's prediction results conform to the physical laws of wireless channels, this invention also includes an attention-coherence consistency verification step. The verification process and result analysis are as follows: Figure 5 As shown. Based on Clarke-style channel coherence theory, three verifiable hypotheses are proposed, and attention-coherence consistency analysis is carried out: The first type is time-scale alignment verification, which calculates the effective backtracking length L of the model's attention mechanism. eff The first type of verification verifies the negative correlation between this length and channel Doppler spread, i.e., the higher the carrier frequency and user mobility speed, the greater the Doppler spread, and the shorter the effective attention backtracking length should be, ensuring that the model's attention mechanism can adapt to the temporal evolution characteristics of the channel; the second type is the hysteresis contour shape alignment verification, which aggregates the attention hysteresis distributions of the model layers-wise in each physical domain and calculates the median hysteresis L of the distribution. 50 Compared to interquartile range (IQR), the model's attention distribution in the high Doppler domain is more concentrated, and the median lag decreases monotonically with increasing carrier frequency and user movement speed, ensuring a high degree of match between the model's attention distribution and the channel's coherence characteristics. The third type of verification is context truncation consistency verification. Gradient context truncation is performed on samples from different physical domains, gradually decreasing from the longest to the shortest. The change in predicted NMSE after each truncation is calculated, verifying that context truncation monotonically reduces the model's prediction performance. Furthermore, the low Doppler scenario in the long coherence domain has a higher tolerance for context truncation, ensuring that the model's attention mechanism effectively learns the physical laws of the channel, rather than engaging in meaningless feature fitting. Through these three types of verification, the model's prediction results possess both high accuracy and robustness, and deeply conform to the physical laws of the wireless channel, significantly improving the reliability and stability of the model's deployment in real-world wireless communication scenarios.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for autoregressive prediction of physically conditional channel state information based on a large language model, characterized in that, include: S1. Perform continuous representation processing on the original channel state information matrix collected by the wireless channel, and convert the complex-valued channel state information data into continuous channel state information embedding features that are adapted to the pre-trained large language model sequence modeling. S2. Extract physical metadata, convert it into a prompt token through the physical conditional prompt coding module, and concatenate it to the front end of the sequence of the continuous channel state information embedding features to form a complete physical conditional input sequence. S3. The physical conditional input sequence is fed into a pre-trained large language model to complete autoregressive modeling and output the predicted channel state information matrix. The pre-trained large language model adopts an end-to-end joint training method, uses a combined loss function consisting of mean square error loss and complex angle consistency loss, masks the loss calculation of physical cue tokens, and only optimizes the loss for channel state information sequence features. S4. Introduce a difficulty-aware course training strategy, divide training samples into difficulty levels according to physical metadata, and optimize the pre-trained large language model training process by adopting a phased dynamic sampling method. S5. Using the pre-trained large language model that has been trained, combined with historical channel state information data and physical metadata from real-world scenarios, autoregressive prediction of channel state information at future moments is completed.

2. The method for autoregressive prediction of physically conditional channel state information based on a large language model according to claim 1, characterized in that, The continuous representation processing in S1 specifically involves: decomposing the complex-valued channel state information matrix into real and imaginary parts and then concatenating them into vectors; performing zero-mean-unit-variance normalization on the concatenated data to obtain a real-valued feature vector of uniform scale; projecting this real-valued feature vector onto the hidden dimension space of the pre-trained large language model through a lightweight linear mapping layer to generate continuous channel state information embedding features; and having the learnable parameters of the lightweight linear mapping layer participate in end-to-end joint training.

3. The method for autoregressive prediction of physically conditional channel state information based on a large language model according to claim 1, characterized in that, The physical metadata in S2 includes three core parameters: carrier frequency, propagation scenario type, and user movement speed. The working process of the physical conditional cue encoding module is as follows: after normalizing the carrier frequency and user movement speed respectively, they are fused into Doppler proxy features, and Doppler cue tokens are generated through multi-scale Fourier feature encoding. Input the propagation scenario type into the learnable scenario embedding table and convert it into a scenario category cue token; The Doppler cue token and the scene category cue token are concatenated in a fixed order to form a physical conditional prefix, which is then concatenated with a sequence of continuous channel state information embedding features to form a physical conditional input sequence.

4. The autoregressive prediction method for physically conditional channel state information based on a large language model according to claim 3, characterized in that, The Doppler proxy feature is obtained by multiplying the carrier frequency normalization result with the user movement speed normalization result. The frequency coefficient scale of the multi-scale Fourier feature encoding includes {2,5,10,25,50}, and the number of generated Doppler cue tokens is 5.

5. The method for autoregressive prediction of physically conditional channel state information based on a large language model according to claim 1, characterized in that, The autoregressive modeling in S3 specifically involves: using a general pre-trained large language model backbone, relying on its long sequence modeling capability to generate channel state information embedding features for multiple future time steps through point-by-point autoregression; mapping the channel state information embedding features back to the physical domain through a lightweight linear regression head, and then restoring them to a prediction matrix consistent with the original channel state information format and scale through dimensional reshaping and inverse normalization processing.

6. The autoregressive prediction method for physically conditional channel state information based on a large language model according to claim 1, characterized in that, The combined loss function L in S3 pred Satisfy: L pred =λ1L MSE +λ2L angle Where λ1 and λ2 are weighting coefficients, and λ1 + λ2 = 1; L MSE L represents the complex mean square error, used to constrain the accuracy of CSI amplitude reconstruction. angle This is the phase angle error, used to ensure that the predicted CSI is aligned with the complex phase of the true CSI.

7. The method for autoregressive prediction of physically conditional channel state information based on a large language model according to claim 1, characterized in that, The specific training strategy for the difficulty-aware course in S4 is as follows: the comprehensive difficulty score of the samples is calculated by weighting the carrier frequency rarity, user mobility, and scene propagation complexity, and the training samples are divided into three levels: easy, medium, and hard; dynamic sampling is carried out in three stages: in the early stage of training, easy samples are the main focus, and a small number of medium and hard samples are added. During the mid-training phase, gradually increase the sampling ratio of medium and difficult samples; In the later stages of training, a near-uniform sampling method is adopted with a slight bias towards difficult samples; in the early stages of training, an auxiliary scheduling strategy that prioritizes sampling samples with long context windows is simultaneously adopted.

8. The autoregressive prediction method for physically conditional channel state information based on a large language model according to claim 7, characterized in that, The scores for each dimension of the overall difficulty score of the sample were normalized to the interval [0,1].

9. The autoregressive prediction method for physically conditional channel state information based on a large language model according to claim 1, characterized in that, The autoregressive prediction process of S5 is as follows: collect historical channel state information data and corresponding physical metadata of the actual scenario, complete the continuous representation and physical conditional input sequence construction according to the steps of S1 and S2; generate future multi-step channel state information embedding features from the backbone of the general pre-trained large language model through autoregression, and after linear regression head mapping, dimension reshaping and inverse normalization, output the channel state information prediction result that meets the requirements of wireless communication system.

Images