Wireless channel state deep learning prediction method and system
By using a time-frequency-space three-branch parallel feature extraction network and Doppler adaptive model switching, the problem of insufficient channel prediction accuracy in high-speed mobile scenarios is solved, achieving high-precision channel prediction and improving system performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-07
Smart Images
Figure CN122159988B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication channel estimation and prediction technology, specifically relating to a method and system for wireless channel state prediction based on deep learning. Background Technology
[0002] In modern wireless communication systems, accurate acquisition of channel state information is crucial for achieving efficient adaptive transmission. Multiple-input multiple-output (MIMO) orthogonal frequency division multiplexing (OFDM) systems have become a core physical layer technology for fifth-generation (5G) mobile communications and their evolution, significantly improving spectral efficiency and system capacity by utilizing channel characteristics across spatial, frequency, and temporal dimensions. However, in high-speed mobile scenarios, wireless channels exhibit rapidly time-varying characteristics, significantly shortening channel coherence time. This causes the channel state information acquired by the base station to become outdated by the time it is used for downlink transmission; this phenomenon is known as the channel aging problem. Channel aging directly leads to precoding gain loss, decreased interference suppression capability, and user scheduling decision failure, severely restricting the performance of practical systems in high-mobility scenarios.
[0003] To address the information staleness problem caused by time-varying channels, channel prediction techniques have emerged. Traditional channel prediction methods are mainly based on linear prediction algorithms such as autoregressive models, Wiener filters, and Kalman filters. These methods assume that the time correlation of the channel can be accurately described by a linear model. However, real-world wireless channels are affected by complex factors such as multipath propagation, scatterer motion, and carrier frequency offset, exhibiting significant nonlinear characteristics. Linear prediction methods are clearly insufficient in predicting accuracy under rapidly changing channel conditions. Furthermore, traditional methods typically only utilize time-domain correlation for prediction, neglecting frequency-selective fading correlation between subcarriers and spatial correlation between multiple antennas in the frequency domain, failing to fully exploit the multidimensional structural features inherent in the channel state information.
[0004] Chinese patent CN108566255A discloses a channel prediction method for time-correlated MIMO systems based on multi-task learning. This method treats channel state information from different transmit / receive antenna pairs as different learning tasks, establishes a prediction model using a multi-task least-squares support vector machine, and utilizes the inherent relationships between antenna pairs to improve prediction accuracy. While this approach considers the correlation between multiple antennas, its technical approach has the following limitations: First, the method uses a support vector machine as the core predictor, essentially remaining a shallow machine learning method, which limits its feature extraction and generalization performance when facing high-dimensional, highly nonlinear channel variations. Second, the method only performs sequential prediction of channel state information in the time dimension, without involving correlation modeling between frequency-domain subcarriers. Third, the method utilizes the correlation between multiple antennas by treating each antenna pair as a parallel task, without establishing an explicit graph model of the antenna topology. Fourth, the method does not consider model adaptation under different moving speeds and scattering environments, lacking a mechanism for dynamically adjusting the prediction model according to the channel's time-varying rate.
[0005] In recent years, deep learning technology has demonstrated powerful feature extraction and nonlinear modeling capabilities in the physical layer of wireless communication. Existing research has applied convolutional neural networks and recurrent neural networks to channel estimation and prediction tasks, achieving prediction accuracy superior to traditional methods. However, most existing deep learning channel prediction schemes only focus on joint modeling of the time domain or time-frequency dimensions, and have not yet systematically and effectively fused time-domain sequence features, frequency-domain selective fading features, and spatial antenna topology features. Furthermore, how to dynamically switch network models according to changes in the channel environment to balance prediction accuracy and computational cost remains a pressing technical challenge.
[0006] Therefore, designing a deep learning channel prediction scheme that can simultaneously capture multi-scale features of the channel in the time, frequency, and spatial domains, and possess environmental adaptability, to effectively overcome the system performance degradation caused by channel aging in high-speed mobile scenarios, is a significant challenge for those skilled in the art. Specifically, this scheme needs to ensure high prediction accuracy while maintaining controllable computational complexity, enabling low-latency online inference on the hardware platform of a practical communication system to meet the latency constraints of real-time communication. Summary of the Invention
[0007] To address the aforementioned technical issues, this invention provides a deep learning-based method for predicting wireless channel states. This method constructs a channel state tensor sequence and employs a time-frequency-space three-branch parallel feature extraction network to jointly model the channel's frequency-selective fading characteristics, time-varying dynamic characteristics, and multi-antenna spatial correlation structural characteristics. After adaptive weighted fusion, the model is input into a multi-step prediction head network to generate multi-step channel state prediction values. These multi-step channel state prediction values include channel state prediction values for multiple future time slots. A Doppler adaptive model switching mechanism is designed to dynamically adjust network parameter configurations based on the current channel time-varying rate. Finally, the prediction results are used to perform precoding calculations and resource scheduling in advance, effectively compensating for channel feedback delay.
[0008] Specifically, the deep learning prediction method for wireless channel state provided by this invention includes the following steps: a channel state tensor construction step, a time-frequency-space three-branch feature extraction step, an adaptive weighted feature fusion step, a multi-step channel state prediction step, a Doppler adaptive model switching step, and a prediction-driven precoding and resource scheduling step. Specifically, the channel state tensor construction step performs frequency domain transformation processing on the received pilot signal, organizing the channel frequency response into a tensor sequence according to three dimensions: time slot, antenna, and subcarrier; the time-frequency-space three-branch feature extraction step extracts frequency domain, time domain, and spatial domain features respectively through a one-dimensional causal convolutional network, a long short-term memory network, and a graph attention network; the adaptive weighted feature fusion step generates adaptive weights for each branch through a gated fusion network and performs weighted fusion; the multi-step prediction step generates multi-time-slot prediction values using an autoregressive decoding method; the Doppler adaptive model switching step selects matching network parameter configurations based on the Doppler frequency shift value and uses prediction error feedback to achieve online parameter fine-tuning; and the precoding and resource scheduling step pre-calculates the precoding matrix and resource allocation scheme using the prediction values.
[0009] In the technical solution of this invention, a deeply coupled closed-loop collaborative architecture is formed among the steps: the output of the previous step directly serves as the key input of the next step, while the prediction error feedback mechanism in the Doppler adaptive model switching step enables the evaluation results of the subsequent step to adjust the parameters of the preceding feature extraction network in reverse, achieving end-to-end closed-loop optimization. The synergistic effect of the three feature extraction branches makes the joint modeling effect far exceed the sum of the effects of single-dimensional feature extraction, forming a significant synergistic gain.
[0010] This invention also provides a deep learning prediction system for wireless channel states. This system includes a channel state tensor construction module, a time-frequency-space three-branch feature extraction module, an adaptive weighted feature fusion module, a multi-step channel state prediction module, a Doppler adaptive model switching module, and a precoding and resource scheduling module. Each module corresponds to a step in the aforementioned method. Specifically, the channel state tensor construction module converts the original pilot signal into a structured multi-dimensional tensor input; the time-frequency-space three-branch feature extraction module comprises three parallel sub-modules in the frequency, time, and spatial domains, respectively employing a one-dimensional causal convolutional network, a long short-term memory network, and a graph attention network for feature extraction; the adaptive weighted feature fusion module uses a gated fusion network to dynamically weight and combine the features from the three branches; the multi-step channel state prediction module generates multi-slot prediction values using autoregressive decoding; the Doppler adaptive model switching module dynamically switches and fine-tunes model parameters online based on Doppler frequency shift estimation; and the precoding and resource scheduling module converts the prediction results into a precoding matrix and scheduling decisions.
[0011] Compared with existing technologies, the advantages of this invention are as follows: First, by fully mining the multi-scale structural features of the channel in three dimensions through a time-frequency-space three-branch parallel feature extraction network, the channel prediction accuracy is significantly improved; Second, by achieving an optimal balance between accuracy and computational overhead under different moving speed scenarios through a Doppler adaptive model switching mechanism; Third, by enhancing the system's adaptability to environmental changes through an online fine-tuning mechanism for prediction error feedback; Fourth, by effectively compensating for channel feedback delay through prediction-driven precoding and resource scheduling, the system throughput is improved. Attached Figure Description
[0012] Figure 1 This is a flowchart illustrating the deep learning prediction method for wireless channel state in an embodiment of the present invention.
[0013] Figure 2 This is a schematic diagram of the architecture of the deep learning prediction system for wireless channel state in an embodiment of the present invention. Detailed Implementation
[0014] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0015] Reference Figure 1This invention provides a deep learning method for predicting wireless channel states, which is applied to wireless communication base stations or receiving terminals equipped with multiple antenna arrays. This embodiment uses an orthogonal frequency division multiplexing (OFDM) multiple-input multiple-output (MIMO) communication system as a typical application scenario, wherein the base station is equipped with... Root antenna port, number of system subcarriers Subcarrier spacing The system operates in the 3.5GHz frequency band. In a typical urban macrocell deployment scenario, user terminals within the serving cell may be in different motion states, with movement speeds ranging from approximately 3 km / h (walking speed) to approximately 120 km / h (highway driving speed), corresponding to a maximum Doppler frequency shift of approximately 39 Hz to 389 Hz. The specific implementation methods for each step are described in detail below.
[0016] Step S1: Channel State Tensor Construction. In this step, the system first preprocesses the uplink pilot signal collected by the receiver and extracts the channel frequency response. Specifically, the user terminal transmits known pilot symbols at designated time-frequency resource locations according to a predefined pilot pattern, and the base station receiver samples the received signal at the corresponding time-frequency resource locations. Preferably, this embodiment adopts a comb-shaped pilot arrangement, that is, within each orthogonal frequency division multiplexing symbol, the pilot signal is evenly distributed on the subcarriers, and the pilot interval is one pilot inserted every four subcarriers.
[0017] For the received signal at each pilot position, the system obtains an initial estimate of the channel frequency response at that position by multiplying the received signal by the conjugate of a known pilot symbol. In one embodiment of the invention, let the first... The first time slot, the first The antenna port, the first The received signal on each subcarrier is The corresponding known pilot symbol is The initial estimate of the channel frequency response at that location is: ,in: For the first The first time slot, the first The antenna port, the first The least squares estimate of the channel frequency response at each subcarrier position, which is a complex value; These are the received signal sample values at the corresponding locations, with units consistent with the signal amplitude. For the known pilot symbols at the corresponding positions, satisfy the constant mode constraint. ; Subscript This represents the time slot index, with a value range of [value range missing]. , The length of the historical observation window; subscript This represents the antenna port index, with a value range of [value range missing]. ; Subscript This represents the subcarrier index, with a value range of [value range missing]. .
[0018] After obtaining the initial channel estimate at the pilot position, the channel frequency response at non-pilot positions needs to be interpolated and recovered. This embodiment uses a linear interpolation method to complete the channel response of subcarriers between adjacent pilots along the frequency domain direction. Preferably, for scenarios with low signal-to-noise ratios, pilot denoising processing can be performed on the initial estimate at the pilot position before interpolation. In one embodiment of the present invention, the pilot denoising processing employs an adaptive low-pass filtering method based on the channel frequency response power spectrum distribution. Specifically, an inverse discrete Fourier transform is performed on the initial estimate on all subcarriers to obtain a time-delay domain representation, retaining the first subcarrier with power exceeding the noise floor. After setting one tap and the remaining taps to zero, the frequency domain is transformed back. The number of taps corresponding to the length of the loop prefix is taken as a value in this embodiment. The filter cutoff frequency for this denoising process is dynamically adjusted based on the signal-to-noise ratio (SNR) estimate. When the SNR is higher than 20 dB, all taps are retained to maintain frequency resolution; when the SNR is lower than 10 dB, only the first tap is retained. A tap is used to enhance noise suppression.
[0019] After the above processing, the channel frequency response data for each time slot is arranged into a two-dimensional matrix according to the antenna port and subcarrier dimensions. Furthermore, the real and imaginary parts of the channel frequency response are stored separately as two independent channels. Thus, the channel state information of a single time slot is represented as a two-dimensional matrix. The three-dimensional tensor. Finally, the continuous The channel state tensors of each time slot are stacked along the time axis to form a dimension of The historical observation window tensor. In this embodiment, the historical observation window length... The value of is determined based on the channel coherence time, ranging from 8 to 64 time slots. Smaller values are suitable for high-speed scenarios with rapidly changing channels to reduce computational latency, while larger values are suitable for low-speed scenarios with slowly changing channels to fully utilize historical information. A preferred value is . This value ensures that the observation window covers approximately four coherent time intervals, which effectively captures the temporal correlation characteristics of the channel without introducing excessive historical information redundancy due to an overly long window. Thus, step S1 completes the transformation from the raw received signal to a structured channel state tensor sequence, and its output tensor sequence serves as the input to the three-branch feature extraction network in subsequent step S2.
[0020] Step S2: Time-Frequency-Spatial Three-Branch Feature Extraction. This step is the core innovation of the entire prediction method. It uses three parallel feature extraction branches to perform deep feature mining on the channel state tensor from the frequency, time, and spatial domains respectively. The motivation for this three-branch parallel architecture is that the frequency-selective fading characteristics, time-varying dynamic characteristics, and multi-antenna spatial correlation structure characteristics of the wireless channel exhibit different data distribution patterns and statistical properties. Using a targeted network structure can more effectively extract feature representations from each dimension. The specific implementation of each of the three branches is explained below.
[0021] The goal of frequency domain branching is to capture the frequency-selective fading correlation between subcarriers. In wideband orthogonal frequency division multiplexing (OFDM) systems, multipath propagation causes channel fading on different subcarriers to exhibit frequency-selective characteristics, with significant correlations in the channel responses between adjacent subcarriers. The bandwidth of this correlation is determined by the channel's coherence bandwidth. This invention employs a one-dimensional causal convolutional network to perform sliding window operations along the subcarrier dimension to extract this frequency correlation feature.
[0022] Specifically, the input to the frequency domain branch is a slice of the channel state tensor in the frequency domain dimension, that is, for a given time slot and antenna port The channel frequency response vectors of all subcarriers on the antenna in the given time slot are extracted. In this invention, the kernel length of each causal convolutional layer is selected from 3 to 7, where a smaller kernel length is suitable for dense multipath scenarios with narrow coherence bandwidth to achieve fine frequency resolution, and a larger kernel length is suitable for sparse multipath scenarios with wide coherence bandwidth to expand the receptive field coverage. In one embodiment of this invention, the one-dimensional causal convolutional network includes three causal convolutional layers, configured as follows: the kernel length of the first causal convolutional layer is 5, the number of input channels is 2 (real and imaginary parts), the number of output channels is 64, and the dilation coefficient is 1; the kernel length of the second causal convolutional layer is 5, the number of input channels is 64, the number of output channels is 128, and the dilation coefficient is 2; the kernel length of the third causal convolutional layer is 3, the number of input channels is 128, the number of output channels is 256, and the dilation coefficient is 4. Each causal convolutional layer is followed by a batch normalization layer and a leaky linear rectified activation function, where the negative half-axis slope of the leaky linear rectified activation function is set to 0.01.
[0023] A key characteristic of causal convolution is that its padding method only performs zero padding on the left side of the input sequence, ensuring that the first... The output at each subcarrier position depends only on the index not exceeding [a certain value]. The subcarrier input, a characteristic that avoids future information leakage problems. Due to the doubling of the dilation coefficient layer by layer, the receptive field of the 3-layer causal convolutional network covers... Each subcarrier, with a subcarrier spacing of 30kHz, corresponds to a frequency range of 630kHz. This range is approximately 2 to 3 times the channel coherence bandwidth in a typical urban macrocell scenario, effectively capturing the relevant structure of frequency-selective fading. Finally, the frequency domain branch performs a global average pooling operation on the output of layer 3 along the subcarrier dimension, obtaining a frequency domain feature vector of dimension 256. .
[0024] The goal of the time-domain branch is to model the dynamic characteristics of channel state information over time. The time-varying characteristics of a wireless channel originate from the relative motion of the transmitter, receiver, and scatterer, specifically manifested as changes in the amplitude and phase of the channel impulse response over time, with the rate of change directly related to the Doppler shift. The time-domain branch employs a two-layer long short-term memory network to model the channel state information sequence within the historical observation window along the time slot dimension.
[0025] Specifically, the input to the time-domain branch is the sequence of the channel state tensor in the time dimension, that is, for a given antenna port... and subcarriers Extract continuous The channel frequency response at this location in each time slot constitutes a time series. In one embodiment of the present invention, the Long Short-Term Memory (LSTM) network is configured as follows: the input dimension of the first LSM layer is 2 (real and imaginary parts), the hidden layer dimension is 128, and bidirectional processing mode is enabled; the input dimension of the second LSM layer is 256 (after bidirectional concatenation), the hidden layer dimension is 128, and bidirectional processing mode is also enabled. A random deactivation layer with a dropout rate of 0.2 is set between the two layers to prevent overfitting.
[0026] Long Short-Term Memory (LSTM) networks can effectively learn long-term dependencies in time series through their gating mechanism. Internally, the forget gate determines the proportion of historical information retained, the input gate controls the injection of new information, and the output gate adjusts the feature output at the current time step. In one embodiment of this invention, the LTM network in the... The update process of each time slot can be described as follows:
[0027] ,
[0028] ,
[0029] ,
[0030] ,
[0031] ,
[0032] ,
[0033] in: For the first The forget gate output vector for each time slot has a dimension of 128, which is the same as the hidden layer dimension, and the value range is (0,1). The input gate output vector; The candidate memory unit vector; This is the state vector of the memory cell; The output gate outputs a vector; Output vectors for the hidden states; For the first The input vector for each time slot; This represents the sigmoid activation function; Represents the hyperbolic tangent activation function; This represents element-wise multiplication. These are the weight matrices for the forget gate, input gate, candidate memory unit, and output gate, respectively, each with dimension [missing information]. , where 128 is the hidden layer dimension and 2 is the input dimension; These are the corresponding bias vectors, each with a dimension of 128; This indicates a vector concatenation operation.
[0034] The time-domain branch takes the last time slot. The concatenation of the bidirectional hidden states is used as the temporal feature vector to obtain... .
[0035] The goal of the spatial branch is to learn the spatial correlation structure between multiple antenna ports. In multi-antenna systems, spatial correlations exist between channel responses at different antenna ports due to the physical arrangement of the antenna array and the influence of the scattering environment. This invention innovatively models the multi-antenna system as a graph structure, where each antenna port is considered a node in the graph. The spatial correlation between antennas is characterized by edge weights, and a graph attention network is used to adaptively learn the information aggregation between nodes.
[0036] Specifically, the spatial branch first constructs the antenna diagram. , where the set of nodes Includes all One antenna port, edge set The construction rule is as follows: when the physical distance between two antenna ports in the array is less than a preset threshold (preferably set to 5 times the wavelength, i.e., approximately 0.43m), a connection edge is established between the two nodes. For a uniform linear array configuration, each antenna node is connected to its two neighboring nodes before and after it. The initial feature of each node is the channel frequency response vector of the current time slot on that antenna port, with a dimension of... .
[0037] In one embodiment of the present invention, the graph attention network comprises two graph attention layers, each employing a four-head multi-head attention mechanism. In each layer, nodes... Its neighboring nodes The attention coefficient is calculated as follows:
[0038] ,
[0039] ,
[0040] in: For nodes For nodes The original attention score is a scalar value; The final attention coefficients after processing with the normalized exponential function satisfy... ; For nodes The input feature vector of the first layer The second floor ; For the shared linear transformation weight matrix, where The transformed feature dimensions are in the first layer. (Each attention head), in the second layer ; These are the attention vector parameters; This represents a vector concatenation operation; Represents a node The set of neighboring nodes; It is a linear rectification activation function with leakage, and the slope of the negative half axis is 0.2.
[0041] node The updated features are obtained by attention-weighted aggregation of the features of neighboring nodes:
[0042] ,
[0043] in: For nodes The updated feature vector; This is the exponential linear unit activation function. For multi-head attention, the outputs of each attention head are concatenated along the feature dimension, with the first layer's output dimension being... The output dimension of the second layer is also... Finally, the spatial branch performs global average pooling on the output feature vectors of all nodes to obtain the spatial feature vectors. .
[0044] It is worth noting that the key advantage of graph attention networks lies in the fact that their attention coefficients are learned in a data-driven manner, enabling them to adaptively discover spatial correlation patterns between antennas, rather than relying on pre-assumed spatial correlation models. In real-world scattering environments, the correlation patterns between antennas may exhibit complex, non-uniform structures due to variations in scatterer distribution. Graph attention mechanisms can flexibly capture these dynamically changing correlation structures.
[0045] Step S3: Adaptive weighted feature fusion. After completing the feature extraction of the three branches, this step will fuse the frequency domain feature vector. Temporal feature vectors and spatial eigenvectors Adaptive weighted fusion is performed. Unlike simple splicing or fixed-weight addition, this invention designs a gated fusion network to dynamically generate the weight coefficients of each branch, enabling the system to automatically adjust the contribution ratio of each dimension feature according to the characteristics of the current channel environment.
[0046] Specifically, the feature vectors of the three branches are first concatenated along the feature dimension to obtain the joint feature vector. Subsequently, the joint feature vector is input into the gated fusion network to generate weight coefficients. In one embodiment of the present invention, the gated fusion network comprises a fully connected layer and a normalized exponential function output layer, and its weight generation process is described as follows:
[0047] ,
[0048] in: For the weight vector, The weight coefficients corresponding to the frequency domain branch, time domain branch, and spatial domain branch, respectively, satisfy the following conditions: And the value range of each component is (0,1); This is the weight matrix of the fully connected layer; It is the bias vector; This represents the normalized exponential function, ensuring that the sum of the output weights is 1.
[0049] The fused feature vectors are obtained by weighted summation:
[0050] ,
[0051] in: This results in the final fused feature vector. The technical advantage of this adaptive fusion mechanism lies in its ability to optimize the weights of the frequency domain branch in scenarios with strong channel frequency selectivity (such as multipath-rich urban environments). It will be automatically boosted; in high-speed mobile scenarios where the channel changes rapidly over time, the weight of the time-domain branch will be increased. It will increase; in line-of-sight propagation scenarios where there is significant spatial correlation between antenna arrays, the weight of the spatial branch will increase. This will enhance the results accordingly. This dynamic weight adjustment ensures that the fused feature vector always contains the most predictive information dimension.
[0052] Preferably, an intermediate hidden layer of dimension 128 can be added before the fully connected layer, and a nonlinear transformation is performed using a leaky linear rectified activation function to enhance the feature discrimination capability of the gated network. In this embodiment, after training, in a typical urban macrocell scenario, the average weight of the three branches is approximately , , This indicates that temporal features contribute the most in this scenario.
[0053] Furthermore, in one embodiment of the present invention, a feature dimension alignment mechanism is introduced to ensure that the three branches have a consistent feature representation space before weighted fusion. Specifically, a linear projection layer with dimension 256 is set at the output of each branch to uniformly map the feature vectors output by each branch to the same embedding space. This alignment operation enables the gated fusion network to perform weight allocation in a semantically consistent feature space, avoiding the weight shift problem caused by excessive differences in the feature distribution of each branch. In addition, the fused feature vector undergoes a layer normalization operation and a residual connection before being input into the multi-step prediction head. That is, the fused vector is added to the simple average of the features of the three branches and then normalized. This residual connection design can preserve the original information of each branch and alleviate the gradient vanishing problem in deep networks.
[0054] Step S4: Multi-step channel state prediction. This step utilizes fused feature vectors. As input, a multi-step prediction head network sequentially generates multi-step channel state prediction values using an autoregressive decoding method. These multi-step channel state prediction values include channel state prediction values for multiple future time slots. In one embodiment of the invention, the multi-step prediction head network employs a Transformer decoder structure to fully utilize the global perception capability of the attention mechanism for historical feature information.
[0055] Specifically, the multi-step prediction head network comprises three Transformer decoder layers, each containing one multi-head self-attention sub-layer and one feedforward neural network sub-layer. The multi-head self-attention sub-layer uses eight attention heads, each with a key-value-query dimension of 32, for a total dimension of 256. The feedforward neural network sub-layer contains two fully connected layers with an intermediate dimension expansion of 512, using Gaussian error linear units as the activation function. Each sub-layer is configured with residual connections and layer normalization operations.
[0056] In a multi-step forecasting process, suppose we need to predict the future. The channel state of each time slot, in this embodiment The prediction process employs an autoregressive approach, proceeding sequentially: First, the fused feature vectors are... As the initial input to the decoder, after processing by the Transformer decoder layer, the 256-dimensional features are mapped to a linear projection layer. The channel state prediction value of dimension 1 is used to obtain the prediction result of the first future time slot. Subsequently, After feature encoding, the result is concatenated with the fused feature vector and used as input for the next prediction step, generating the prediction result for the second future time slot. And so on, until all are generated. The prediction results for each time slot.
[0057] During the training phase, this embodiment employs a teacher-forced strategy, using real channel state values instead of the previous step's predictions as the input for the current step to accelerate model convergence and improve training stability. During the inference phase, an autoregressive generation strategy is used, employing the actual predicted outputs as the input for the next step. The loss function for model training is the mean squared error loss.
[0058] ,
[0059] in: The training loss value is a non-negative real number, with units consistent with the square of the channel response amplitude; For the first The first time slot, the first The antenna port, the first Channel state prediction values at each subcarrier position; This represents the actual channel state value at the corresponding location; The prediction step is 2 to 8. Smaller prediction steps are suitable for scenarios with high prediction accuracy requirements, while larger prediction steps are suitable for scenarios that require longer look-ahead time to support scheduling decisions. In this embodiment, the value is 4. This represents the square operation modulo a complex number. The training process uses the Adam optimizer, with an initial learning rate set to... The weight decay coefficient is The batch size is 64, and the total number of training rounds is 200. Preferably, this embodiment also employs a cosine annealing learning rate scheduling strategy, gradually decaying the learning rate from the initial value according to a cosine function curve to... This allows for more precise parameter tuning in the later stages of training. The training dataset is constructed as follows: a large number of channel implementation samples are generated under different channel model parameter configurations. For each parameter configuration, 5000 time-series samples are generated, and each sample contains... Channel state information for consecutive time slots, of which the first One time slot is used as input, then... Each time slot was used as the prediction target. To enhance the model's generalization ability, the training set included channel samples under different moving speeds, different scattering environments, and different signal-to-noise ratios.
[0060] In one embodiment of the present invention, a prediction confidence evaluation mechanism is also introduced during the multi-step prediction process. Specifically, in addition to outputting the predicted channel state value, the multi-step prediction head network also outputs the prediction variance estimate for each prediction step. This prediction variance reflects the uncertainty of the model regarding its own prediction results, and it typically increases with the number of prediction steps. This variance estimate can be used as a channel quality reliability indicator in subsequent precoding calculations and resource scheduling. For future time slots with excessively large prediction variances, the system can choose to adopt a more conservative transmission strategy to reduce the packet error rate.
[0061] Step S5: Doppler Adaptive Model Switching. This step is the key mechanism for achieving a balance between accuracy and computational overhead in this invention. In practical communication systems, different user terminals may be in completely different states of motion, from stationary to high-speed movement, with corresponding significant differences in channel time-varying rates. If the same deep prediction model is used uniformly for all users, computational resources will be wasted in low-speed scenarios, while in high-speed scenarios, insufficient network capacity may lead to a decrease in prediction accuracy. Therefore, this invention designs a Doppler adaptive model switching mechanism to dynamically select a suitable prediction model configuration based on the current channel time-varying rate.
[0062] First, the system determines the current speed level by estimating the Doppler frequency shift of the received signal. Preferably, this embodiment uses an autocorrelation method based on the pilot signal phase difference to estimate the Doppler frequency shift. Specifically, the time autocorrelation function of the channel frequency response at the same subcarrier position in adjacent time slots is calculated, and the frequency corresponding to the delay when the normalized amplitude of the autocorrelation function first drops below a preset threshold is taken as the estimated value of the Doppler frequency shift.
[0063] This invention divides the movement speed into three levels, and the definitions and corresponding model configurations for each level are as follows:
[0064] Low speed level: when Doppler frequency shift (Approximately the mobile speed in the 3.5GHz band) When the channel changes slowly, a lightweight model configuration is used to reduce computational overhead. In this lightweight configuration, the number of causal convolutional layers in the frequency domain branch is reduced to 2, the long short-term memory network in the time domain branch is simplified to a single layer and the hidden layer dimension is reduced to 64, and the number of graph attention layers in the spatial domain branch is reduced to 1 and the number of attention heads is reduced to 2. The floating-point computation cost of this lightweight configuration is approximately 25% of that of the full model, while the prediction accuracy decreases by only about 2% in low-speed scenarios.
[0065] Medium speed level: when the Doppler frequency shift is (Corresponding movement speed approximately) When the range is within the specified range, the standard model configuration is adopted, namely the complete three-branch network structure described in step S2 above.
[0066] High-speed rating: When Doppler frequency shift (Corresponding movement speed approximately) When the channel changes rapidly, an enhanced deep model configuration is activated to ensure prediction accuracy. In this enhanced configuration, the number of causal convolutional layers in the frequency domain branch is increased to 4, and the maximum dilation coefficient is increased to 8; the long short-term memory network in the time domain branch is increased to 3 layers, and the hidden layer dimension is increased to 256; and the number of graph attention layers in the spatial domain branch is increased to 3 layers, with each layer containing 8 attention heads. Furthermore, the enhanced configuration introduces additional cross-attention sub-layers in the Transformer decoder, enabling the decoder to more fully utilize encoded feature information.
[0067] It is worth emphasizing that this step also includes an online parameter fine-tuning mechanism for prediction error feedback, which is a key link in forming a closed-loop synergy in the entire method. Specifically, the system continuously monitors the normalized mean square error between the prediction results and the actual measured values, and accumulates the most recent mean square error using an exponential moving average method. The prediction error for each sample. When the cumulative error exceeds a preset threshold. When the system determines that the current channel environment has changed significantly, it triggers online gradient updates of the model parameters. The learning rate for online fine-tuning is set to 0.01 to 0.1 times the offline training learning rate. A larger rate is used when the channel environment changes drastically to accelerate model adaptation, while a smaller rate is used when the channel environment only experiences slight drift to ensure fine-tuning stability. In this embodiment, the learning rate for online fine-tuning is set to 0.05 times the offline training learning rate, i.e. Each fine-tuning update only updates the parameters of the last two layers of the network (i.e., the fusion network and the prediction head), while freezing the low-level parameters of the feature extraction network to prevent catastrophic forgetting. This online fine-tuning mechanism allows error information from the prediction results to propagate back and influence the behavior of the preceding feature extraction and fusion steps, forming an end-to-end closed-loop optimization structure.
[0068] It is worth further explaining that the speed level switching involved in the Doppler adaptive model switching process has a certain lag effect. To avoid increased system overhead and fluctuations in prediction accuracy caused by frequent model switching near the boundary between two speed levels, this embodiment introduces a hysteresis mechanism. Specifically, the uplink threshold for switching from low speed to medium speed is set to 55Hz, while the downlink threshold for switching from medium speed back to low speed is set to 45Hz. The difference between the two thresholds is the hysteresis window width of 10Hz. Similarly, the uplink threshold between medium speed and high speed is set to 320Hz, the downlink threshold is set to 280Hz, and the hysteresis window width is 40Hz. The actual model parameter switching operation is only triggered when the estimated value of the Doppler frequency shift remains within the judgment range of the new level continuously within a preset time window (preferably 10 consecutive time slots). The switching process adopts a hot switching method, that is, the current model continues to be used for prediction until the new model parameters are loaded, ensuring that there is no prediction interruption during the switching process. The loading time of the model parameters on the hardware platform of this embodiment is about 0.2ms, which is much less than the duration of a time slot, so its impact on the real-time performance of the system is negligible.
[0069] Step S6: Prediction-driven precoding and resource scheduling. This step utilizes the multi-step channel state prediction values output from step S4 to perform precoding matrix calculation and resource block allocation scheduling decisions in advance, thereby effectively compensating for the performance loss caused by channel feedback delay.
[0070] In terms of precoding matrix calculation, this embodiment employs a zero-forcing precoding method based on channel state prediction values. Specifically, it utilizes the first... Channel state prediction matrix for each time slot Calculate the pseudo-inverse matrix as the precoding matrix:
[0071] ,
[0072] in: For the first The precoding matrix for each time slot, The number of users scheduled simultaneously; The predicted downlink channel matrix; This represents the conjugate transpose operation; This represents the matrix inversion operation. After calculation, the precoding matrix is normalized for power.
[0073] In terms of resource scheduling, the system performs forward-looking resource block allocation based on predicted multi-step channel state information. Preferably, this embodiment employs a proportional-fair scheduling algorithm, using the predicted channel capacity of future time slots instead of the instantaneous channel capacity at the current moment as the scheduling criterion. For each user, based on their future... Resource block allocation priority is ranked by the weighted average of the predicted channel capacity over each time slot. The weights decrease exponentially with time slot distance to reflect the characteristic that prediction accuracy decreases with the number of prediction steps. Through this prediction-driven scheduling strategy, the system can allocate resources in advance to users with better channel conditions in the future, thereby improving the average throughput of the system on a macro level.
[0074] In one embodiment of the present invention, the specific implementation process of resource scheduling is as follows: Assume that the system has a total of There are 1 active user in each subframe. The number of allocable resource blocks. For the number of... The user in the first The scheduling priority on each resource block is determined by the system based on the ratio of the predicted channel capacity of the user in the corresponding frequency band of that resource block to its historical average throughput. The predicted channel capacity is calculated based on the channel state prediction value output in step S4, estimated using the Shannon capacity formula. Preferably, for multi-step prediction, the system performs a weighted average of the predicted capacities for each future time slot, where the... The weights of the prediction results are set as follows: Attenuation factor The value is set to 0.85, which gives higher weight to the prediction results of more recent time slots, reflecting the objective law that the prediction accuracy decreases as the number of steps increases.
[0075] Furthermore, in one embodiment of the present invention, the calculation of the precoding matrix also incorporates prediction confidence information. When the prediction variance of a certain time slot exceeds a preset confidence threshold, the system automatically introduces a regularization factor into the precoding calculation to enhance numerical stability. Specifically, the matrix inversion in the precoding formula is replaced with an inversion operation with a regularization term, and the regularization coefficient is proportional to the prediction variance. This mechanism effectively avoids power amplification and beam pointing deviation problems caused by an excessively large condition number of the precoding matrix when the channel prediction uncertainty is high, ensuring the robustness of the system under various channel conditions.
[0076] Reference Figure 2 The present invention also provides a deep learning prediction system for wireless channel states, wherein each functional module of the system corresponds one-to-one with each step in the aforementioned method embodiments, and specifically includes the following modules:
[0077] The channel state tensor construction module, corresponding to step S1 in the method embodiment, is responsible for receiving the uplink pilot signal and performing frequency domain transformation, pilot denoising, and interpolation operations on it. It extracts the channel frequency response of all antenna ports on all subcarriers and organizes the channel frequency response of multiple consecutive time slots into a channel state tensor sequence according to four dimensions: time slot, antenna port, subcarrier, and real / imaginary parts. The output interface of this module directly interfaces with the input interface of the time-frequency-space three-branch feature extraction module, and the output tensor sequence has the following dimensions: The specific dimensional parameters are consistent with those described in the aforementioned method embodiments.
[0078] The time-frequency-space three-branch feature extraction module, corresponding to step S2 in the method embodiment, is the core computational module of the entire system. This module contains three parallel feature extraction sub-modules: the frequency domain feature extraction sub-module uses a one-dimensional causal convolutional network, with its network layer number, kernel length, dilation coefficient, and channel number configuration completely consistent with those described in the aforementioned method embodiment, outputting a 256-dimensional frequency domain feature vector; the time domain feature extraction sub-module uses a bidirectional long short-term memory network, with its layer number, hidden layer dimension, and random deactivation configuration as described previously, outputting a 256-dimensional time domain feature vector; and the spatial domain feature extraction sub-module uses a graph attention network, with its antenna graph construction rules, attention layer number, and multi-head configuration as described previously, outputting a 256-dimensional spatial domain feature vector. These three sub-modules can be deployed on parallel computing hardware to achieve synchronous execution, thereby effectively reducing inference latency.
[0079] The adaptive weighted feature fusion module, corresponding to step S3 in the method embodiment, receives feature vectors output by the three feature extraction sub-modules, dynamically generates weight coefficients through a gated fusion network, and performs a weighted fusion operation. In one embodiment of the present invention, this module can be deployed on the same inference accelerator as the time-frequency-space three-branch feature extraction module to reduce data transmission overhead.
[0080] The multi-step channel state prediction module, corresponding to step S4 in the method embodiment, receives the fused feature vector and generates multi-step channel state prediction values in an autoregressive manner using a Transformer decoder structure. These multi-step channel state prediction values include channel state prediction values for multiple future time slots. The number of prediction steps and network structure configuration of this module are consistent with those described in the aforementioned method embodiment. The prediction results are simultaneously output to the precoding and resource scheduling module and the Doppler adaptive model switching module.
[0081] The Doppler adaptive model switching module, corresponding to step S5 in the method embodiment, is responsible for Doppler frequency shift estimation, motion speed level determination, and model parameter switching control. This module maintains a model parameter library, storing network parameter sets pre-trained for low, medium, and high speed levels. When a change in motion speed level is detected, this module retrieves the corresponding parameter set from the model parameter library and loads it into the time-frequency-space three-branch feature extraction module and the multi-step channel state prediction module. Simultaneously, this module continuously receives the error signal between the predicted value output by the prediction module and the actual measured value, and determines whether online parameter fine-tuning needs to be triggered based on this signal. Preferably, the online fine-tuning process is executed in an independent background computing thread to avoid affecting the real-time performance of the foreground prediction inference.
[0082] The precoding and resource scheduling module, corresponding to step S6 in the method embodiment, receives multi-step channel state prediction values and calculates the precoding matrix and executes resource block allocation scheduling decisions accordingly. The precoding calculation method and scheduling strategy of this module are consistent with those described in the aforementioned method embodiment. The calculated precoding matrix is directly sent to the base station's radio frequency front-end for downlink beamforming processing, and the scheduling result is sent to the media access control layer for resource allocation.
[0083] In the system implementation, the data flow among the six modules forms a collaborative architecture combining single-feedforward and feedback closed-loop. In the feedforward direction, data flows sequentially from the channel state tensor construction module through the three-branch feature extraction module, the adaptive weighted feature fusion module, and the multi-step prediction module, finally reaching the precoding and resource scheduling module. In the feedback direction, the Doppler adaptive model switching module exerts reverse control on the behavior of the feature extraction and prediction modules through two paths: model parameter switching and online fine-tuning. This feedforward-feedback collaborative architecture enables the system to automatically adjust its prediction behavior according to changes in the channel environment, forming an adaptive closed-loop intelligent prediction system.
[0084] From a hardware deployment perspective, in one embodiment of this invention, the channel state tensor construction module and the Doppler adaptive model switching module are deployed on the general-purpose processor of the baseband processing unit, utilizing its flexible control logic to handle tasks such as pilot signal parsing and model switching judgment. The time-frequency-space three-branch feature extraction module, the adaptive weighted feature fusion module, and the multi-step channel state prediction module are deployed on a dedicated neural network inference accelerator to fully utilize its massively parallel computing capabilities for low-latency inference. Preferably, the above three computationally intensive modules are optimized using model quantization technology, quantizing network parameters from 32-bit floating-point quantization to 8-bit integer representation. Under the condition that the inference accuracy loss does not exceed 0.3dB, the inference speed is increased by about 3.5 times, and the model storage space is reduced by about 75%. The precoding and resource scheduling module is deployed on the media access control layer processor, working in conjunction with the existing resource management function module. After receiving the prediction results, this module immediately starts the parallel computing pipeline of the precoding matrix to ensure that all necessary precoding and scheduling calculations are completed before the next transmission time slot arrives.
[0085] In one embodiment of the present invention, the end-to-end inference latency of the entire system is approximately 0.8 ms, of which channel state tensor construction takes approximately 0.1 ms, three-branch feature extraction takes approximately 0.4 ms, feature fusion and multi-step prediction take approximately 0.2 ms, and precoding computation takes approximately 0.1 ms. This latency meets the requirement of less than 1 ms processing latency in fifth-generation mobile communication systems. Regarding storage overhead, the total number of parameters in the standard model configuration is approximately 2.3M, the lightweight configuration is approximately 0.6M, and the enhanced configuration is approximately 5.8M, all within the storage capacity of the base station hardware.
[0086] To verify the effectiveness of the technical solution of this invention, a comparative simulation experiment was conducted under a standardized channel model. The simulation platform used a Python 3.9 programming environment, the deep learning framework used PyTorch 2.0, and the hardware platform was a server configured with an NVIDIA A100 GPU. The channel model adopted the urban macrocell non-line-of-sight propagation scenario model defined in the 3GPP TR38.901 specification, with the carrier frequency set to 3.5 GHz, the system bandwidth to 100 MHz, the subcarrier spacing to 30 kHz, the number of subcarriers to 256, and the antenna configuration to 32 antennas for the base station and 4 antennas for the user terminal.
[0087] Three moving speed scenarios were set up for testing in the experiment: low speed scenario (walking speed 3km / h, corresponding to a Doppler frequency shift of about 11Hz), medium speed scenario (vehicle speed 60km / h, corresponding to a Doppler frequency shift of about 194Hz), and high speed scenario (high-speed rail speed 350km / h, corresponding to a Doppler frequency shift of about 1130Hz, exceeding the system design range but used for extreme testing). 10,000 test samples were generated for each scenario, and the number of predicted steps was set to 4 time slots.
[0088] The method of this invention is compared with the following benchmark schemes: Scheme A is the multi-task learning support vector machine method described in CN108566255A; Scheme B is the traditional linear Wiener filter prediction method; Scheme C is a single-branch time-domain prediction method using only a long short-term memory network; Scheme D is a two-branch time-frequency joint prediction method using a convolutional neural network and a long short-term memory network. The normalized mean square error is used as the evaluation metric.
[0089] In low-speed scenarios, the normalized mean square error (MSE) of the method of this invention is -28.3 dB, while that of scheme A is -21.5 dB, scheme B is -18.2 dB, scheme C is -24.1 dB, and scheme D is -26.0 dB. In medium-speed scenarios, the MSE of the method of this invention is -22.7 dB, while that of scheme A is -14.8 dB, scheme B is -10.3 dB, scheme C is -17.5 dB, and scheme D is -19.6 dB. In high-speed scenarios, the MSE of the method of this invention is -15.2 dB, while that of scheme A is -6.1 dB, scheme B is -3.5 dB, scheme C is -9.8 dB, and scheme D is -12.1 dB.
[0090] The experimental results show that: First, the method of this invention significantly outperforms all comparative schemes in all test scenarios, especially in high-speed scenarios, where the prediction accuracy of the method of this invention is improved by about 9.1 dB compared with the CN108566255A scheme, fully demonstrating the technical advantages of joint time-frequency-space three-branch modeling and Doppler adaptive model switching; Second, compared with scheme C which only uses time-domain features, this invention brings a performance gain of about 3 to 5 dB by introducing frequency-domain and spatial-domain feature branches, verifying the nonlinear gain effect of multi-dimensional feature synergistic fusion; Third, the Doppler adaptive model switching mechanism makes the computational cost of the method of this invention only about 25% of the standard configuration in low-speed scenarios, effectively reducing unnecessary consumption of computational resources, while ensuring that the prediction accuracy does not decrease in high-speed scenarios through enhanced configuration.
[0091] Regarding system throughput, after using the prediction results of this invention for precoding matrix calculation, the average cell throughput in medium-speed scenarios is improved by about 18.5% compared with the traditional scheme based on outdated channel information, and the gap with the ideal channel feedback scheme without channel prediction is reduced to less than 3.2%, proving that the method of this invention can effectively compensate for the system performance loss caused by channel feedback delay.
[0092] To further analyze the contributions of each technical module, ablation experiments were also conducted in this embodiment. After removing the frequency domain branch, the normalized mean square error in the medium-speed scenario degraded from -22.7dB to -20.1dB, with a performance loss of approximately 2.6dB, indicating that frequency domain features are of significant value in capturing frequency-selective fading information. After removing the spatial domain branch, the performance degraded by approximately 1.8dB to -20.9dB, demonstrating that the graph attention mechanism plays a significant role in modeling the spatial correlation of multiple antennas. When the adaptive fusion mechanism was disabled and simple average fusion was used, the performance degraded by approximately 1.2dB, verifying the dynamic weight adjustment effect of the gated fusion network. When the Doppler adaptive model switching mechanism was disabled and the standard configuration was used uniformly, the computational load increased by approximately 300% in the low-speed scenario, but the prediction accuracy only improved by approximately 0.5dB, while the prediction accuracy decreased by approximately 1.8dB in the high-speed scenario, fully demonstrating the technical value of the adaptive switching mechanism in balancing efficiency and accuracy. After disabling the online fine-tuning mechanism, when the channel environment gradually changed during the test, the prediction accuracy degraded by approximately 2.3 dB after about 500 samples. However, after enabling online fine-tuning, this degradation was controlled within 0.6 dB, verifying the contribution of the prediction error feedback loop to environmental adaptability. The above ablation experiment results fully verify the necessity and effectiveness of the key technical modules proposed in this invention for improving overall prediction performance from various technical dimensions.
[0093] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.
Claims
1. A deep learning prediction method for wireless channel states, characterized in that, Includes the following steps: Step S1, Channel State Tensor Construction Step: Perform frequency domain transformation processing on the pilot signal collected by the receiver, extract the channel frequency response of multiple antenna ports on multiple subcarriers, and organize it into a channel state tensor sequence according to the three dimensions of time slot, antenna port and subcarrier. Among them, the channel frequency response of multiple consecutive time slots is stacked along the time axis to form a historical observation window tensor. Step S2, Time-Frequency-Space Three-Branch Feature Extraction Step: The channel state tensor sequence is subjected to parallel feature extraction through a time-frequency-space three-branch feature extraction network. The time-frequency-space three-branch feature extraction network includes a frequency domain branch, a time domain branch, and a spatial domain branch. The frequency domain branch uses a one-dimensional causal convolutional network to perform sliding window operations along the subcarrier dimension to extract frequency-selective fading related features. The time domain branch uses a long short-term memory network to perform sequence modeling along the time slot dimension to extract time-varying dynamic features of the channel. The spatial domain branch uses a graph attention network to treat each antenna port as a graph node and aggregates neighborhood features based on attention coefficients to extract multi-antenna spatial correlation structural features. Step S3, Adaptive weighted feature fusion step: The feature vectors output by the frequency domain branch, the time domain branch and the spatial domain branch are concatenated and then used to generate adaptive weight coefficients for each branch through a gated fusion network. The features of the three branches are then weighted and summed according to the adaptive weight coefficients to obtain the fused feature vector. Step S4, multi-step channel state prediction step: The fused feature vector is input into the multi-step prediction head network, and multi-step channel state prediction values are generated sequentially through autoregressive decoding. The multi-step channel state prediction values include channel state prediction values for multiple future time slots, and the prediction result of each step is used as the input for the next prediction. Step S5, Doppler adaptive model switching step: Determine the current moving speed level based on the Doppler frequency shift value estimated by the receiver, select network parameters matching the current channel time-varying rate from the preset model parameter library according to the moving speed level, load them into the time-frequency-space three-branch feature extraction network and the multi-step prediction head network, and feed back the prediction error between the multi-step channel state prediction value and the actual measurement value to the model parameter library to trigger online parameter fine-tuning.
2. The method according to claim 1, characterized in that, In step S1, the dimension of the historical observation window tensor is ,in The number of historical observation time slots ranges from 8 to 64. Number of antenna ports The number of subcarriers is represented by the number 2 in the last dimension, which indicates the real and imaginary parts of the channel frequency response.
3. The method according to claim 1, characterized in that, In step S2, the one-dimensional causal convolutional network includes at least two causal convolutional layers. The kernel length of each causal convolutional layer is 3 to 7, and the dilation coefficient increases by one layer at a time. A batch normalization layer and a leaky linear rectified activation function are set between each convolutional layer.
4. The method according to claim 1, characterized in that, In step S5, the moving speed level is divided into three levels: low speed level, medium speed level and high speed level. The low speed level corresponds to a Doppler frequency shift of less than 50Hz, the medium speed level corresponds to a Doppler frequency shift between 50Hz and 300Hz, and the high speed level corresponds to a Doppler frequency shift of greater than 300Hz.
5. The method according to claim 1, characterized in that, In step S2, the attention coefficient of node i to neighboring node j in the graph attention network is calculated as follows: the feature vectors of node i and node j are linearly transformed and concatenated, and the original attention score is obtained by passing them through a single-layer feedforward network activated by a leaky linear rectifier. Then, the original attention scores of all neighboring nodes of node i are normalized by a normalized exponential function to obtain the final attention coefficient.
6. The method according to claim 1, characterized in that, In step S3, the gated fusion network includes a fully connected layer and a normalized exponential function output layer. The input of the fully connected layer is a joint vector concatenated from the feature vectors of the three branches. The normalized exponential function output layer generates a weight vector with a dimension of 3 and the sum of each component is 1. The multi-step prediction head network adopts a Transformer decoder structure, including a multi-head self-attention layer and a feedforward neural network layer. The number of prediction steps is 2 to 8 time slots. In the training phase, a teacher-forced strategy is adopted, while in the inference phase, an autoregressive generation strategy is adopted.
7. The method according to claim 6, characterized in that, In step S5, the online parameter fine-tuning uses an exponential moving average method to accumulate the most recent preset number of prediction error samples. When the accumulated error exceeds a preset threshold, the gradient update of the model parameters is triggered, and the update learning rate is set to 0.01 to 0.1 times the offline training learning rate.
8. The method according to claim 7, characterized in that, Step S1 further includes pilot denoising processing of the channel frequency response. The pilot denoising processing adopts an adaptive low-pass filter based on the power spectrum distribution of the channel frequency response, and the filter cutoff frequency is dynamically adjusted according to the signal-to-noise ratio estimate.
9. The method according to claim 8, characterized in that, It also includes step S6, a prediction-driven precoding and resource scheduling step: using the multi-step channel state prediction values to pre-calculate the precoding matrix and execute resource block allocation and scheduling decisions.
10. A wireless channel state deep learning prediction system, used to implement the method of claim 9, characterized in that, include: The channel state tensor construction module is configured to perform frequency domain transformation processing on the pilot signal collected by the receiver, extract the channel frequency response of multiple antenna ports on multiple subcarriers, and organize it into a channel state tensor sequence according to the three dimensions of time slot, antenna port and subcarrier. The time-frequency-space three-branch feature extraction module is configured to extract features from the channel state tensor sequence in parallel through frequency domain branch, time domain branch and spatial domain branch respectively. The frequency domain branch uses a one-dimensional causal convolutional network to extract frequency selective fading related features, the time domain branch uses a long short-term memory network to extract time-varying dynamic features of the channel, and the spatial branch uses a graph attention network to extract multi-antenna spatial correlation structure features. The adaptive weighted feature fusion module is configured to generate adaptive weight coefficients from the feature vectors output by the three branches via a gated fusion network and then perform weighted fusion. A multi-step channel state prediction module is configured to input a fused feature vector into a multi-step prediction head network to generate multi-step channel state prediction values in an autoregressive decoding manner. The multi-step channel state prediction values include channel state prediction values for multiple future time slots. The Doppler adaptive model switching module is configured to determine the moving speed level based on the Doppler frequency shift estimate and select the matching network parameter configuration from the model parameter library, and feed the prediction error back to the model parameter library to trigger online parameter fine-tuning; The precoding and resource scheduling module is configured to pre-calculate the precoding matrix and execute resource block allocation and scheduling decisions using the multi-step channel state prediction values.