A method for transmitting and receiving a signal of a wireless communication system based on an AI model

Abstract

The application relates to the technical field of wireless communication, in particular to a wireless communication system signal sending and receiving method based on an AI model; containing delay coordinate embedding, neural flow tracking, continuous evolution, Riemann decoding and a closed-loop control unit; the system maps pilots to a high-dimensional phase space by using the Takens principle, reconstructs a channel dynamic topology; the core is to calculate a flow field vector by using a manifold derivative network, continuously deduces time by using a neural differential equation, and eliminates evolution deviation by combining a geometric projection correction term, so that future channel states are accurately predicted; a CSI matrix is solved based on the prediction result to optimize beam emission; the application breaks through the limitation of single-variable observation, and realizes reconstruction and perception from discrete data to the whole picture of a high-dimensional nonlinear system.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a method for transmitting and receiving signals in a wireless communication system based on an AI model. Background Technology

[0002] With the evolution of wireless communication technology, high-speed mobile communication scenarios such as high-speed railways and vehicle-to-everything (V2X) are becoming increasingly widespread, and the nonlinear dynamic characteristics of the channel environment are becoming more and more complex. This complexity poses a severe challenge to the stability and reliability of signal transmission, especially in the real-time acquisition and prediction of channel state information. Currently, communication systems typically rely on discrete pilot observation sequences to infer the channel state and use recurrent neural networks or linear interpolation methods based on discrete time steps for channel prediction and tracking. However, in high-speed mobile environments, existing methods are limited by discrete sampling mechanisms, inevitably leading to information loss and accumulated errors. They are also unable to effectively cope with multipath scattering changes caused by rapid Doppler shifts and are prone to channel aging effects. In addition, traditional models often rely on blind parameter fitting, resulting in a large number of model parameters and low computational efficiency. They are unable to achieve both high-precision prediction and low inference latency within a very short coherence time, and thus cannot meet the performance requirements of real-time communication systems.

[0003] Therefore, how to reconstruct the full picture of the channel dynamic system using limited historical observation data, overcome the limitations of discrete sampling, and achieve advanced and accurate prediction and real-time compensation for high-speed time-varying channels has become an urgent problem to be solved in this field. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for transmitting and receiving signals in a wireless communication system based on an AI model. Specifically, the technical solution of this invention includes:

[0005] The delayed coordinate embedding unit serves as the sensing entry point of the system. It receives historical pilot observation sequences from user equipment and uses the Takens embedding principle to map one-dimensional time-domain signals into geometric state points in high-dimensional phase space, thereby generating potential state vectors that characterize the topology of the channel dynamic system.

[0006] The potential state vector is input into the Lagrange neural flow tracing unit. The tangent direction of the geometric state point on the manifold surface is analyzed by the manifold derivative network. The flow field vector describing the velocity and direction of the channel state motion is calculated and passed to the next level as the evolution control parameter.

[0007] The continuous-time neural evolution unit, as a solver of neural ordinary differential equations, receives the potential state vector as the initial position, performs continuous-time integral deduction based on the slope provided by the flow field vector, and calculates the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time during the deduction process. It uses geometric projection correction terms to force the projection of the state that violates physical inertia, thereby outputting the predicted potential state at a future specified time point.

[0008] The Riemann projection decoding unit uses a nonlinear mapping relationship to project the predicted latent state in the high-dimensional manifold space back into the Euclidean space and solves the complex domain channel state information matrix for future time moments.

[0009] The closed-loop signal transmission control unit calculates the optimal downlink precoding matrix based on the complex domain channel state information matrix, and after mapping the service data stream to be transmitted to the optimal downlink precoding matrix, transmits the beam through the antenna array.

[0010] Preferably, the step of mapping a one-dimensional time-domain signal to geometric state points in a high-dimensional phase space using the Takens embedding principle to generate a latent state vector specifically includes:

[0011] Construct a delay vector matrix to map historical pilot observations at different times into components on a high-dimensional coordinate axis, forming a numerical set that can reflect the whole system.

[0012] The numerical set is defined as a potential state vector and directly input as the initial position into the Lagrange neural flow tracing unit to solve the problem that univariate observations cannot reflect the complete state of the channel dynamics system.

[0013] Preferably, the step of analyzing the tangent direction of the geometric state point on the manifold surface through a manifold derivative network and calculating the flow field vector specifically includes:

[0014] Configure the built-in manifold derivative network to fit the geometric tangent space, and receive the latent state vector as input;

[0015] Analyze the evolutionary physical inertia of the current state point on the surface of the phase space manifold, and generate control parameters that represent the evolutionary derivative of the system;

[0016] The control parameter is defined as a flow field vector, which is used to specify the motion trend of the channel state in phase space and serves as the core basis for the continuous-time neural evolution unit to perform integral operations.

[0017] Preferably, the continuous-time integration derivation based on the slope provided by the flow field vector specifically includes:

[0018] Numerical integration algorithms are used to simulate the particle drift process in physical fields;

[0019] The potential state vector is calculated to its new position after moving along the flow field trajectory under the action of the flow field vector, thereby realizing the simulation of the evolution of the channel state in continuous time.

[0020] Preferably, the step of calculating the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time during the deduction process, and using a geometric projection correction term to force projection of states that violate physical inertia, specifically includes:

[0021] During the integral evolution process, the geometric distance between the current evolution trajectory position and the preset low-dimensional manifold surface is monitored in real time to obtain the manifold deviation value;

[0022] If the manifold deviation value exceeds a preset threshold, it is determined that a jump that violates physical inertia has occurred, and a geometric projection correction term is generated.

[0023] The geometric projection correction term is applied directly to the current integral state vector, forcing it back to the surface of the manifold that conforms to physical laws, and this correction process does not trigger the backpropagation of network weights.

[0024] Preferably, the Riemann projection decoding unit calculates the complex domain channel state information matrix for future time moments, specifically including:

[0025] Configure this unit as an output mapping layer to receive the predicted latent states in an abstract high-dimensional manifold space;

[0026] The predicted latent state is translated into physical communication parameters through a trained nonlinear network layer, and the complex domain channel state information matrix corresponding to time t+Δt is output.

[0027] Preferably, the closed-loop signal transmission control unit calculates the optimal downlink precoding matrix, specifically including:

[0028] Based on the predicted complex domain channel state information matrix for future moments, the calculation is performed using singular value decomposition or zero-forcing algorithm;

[0029] Generate the optimal downlink precoding matrix for the next time step to pre-compensate for Doppler shift and channel aging caused by high-speed movement.

[0030] Preferably, the step of mapping the service data stream to be transmitted to the optimal downlink precoding matrix and then transmitting the beam through the antenna array specifically includes:

[0031] Execute formula Matrix operations, where For the business data stream to be sent, The optimal downlink precoding matrix is... The weighted transmitted signal;

[0032] The large-scale antenna array is controlled to transmit the weighted transmission signal, forming a beam pointing towards the user equipment.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This method utilizes delayed coordinate embedding units based on the embedding principle to map the one-dimensional historical pilot observation sequence of user equipment into geometric state points in a high-dimensional phase space. This mechanism effectively solves the technical problem that single-variable observations in traditional methods cannot reflect the full picture of high-dimensional nonlinear systems. It can reconstruct the potential state vector representing the complete topological structure of the channel dynamic system from limited discrete data. It not only obtains the current amplitude information, but also embeds the higher-order rate of change and acceleration characteristics of the signal, thereby significantly improving the system's ability to perceive complex channel environments.

[0035] 2. This method employs Lagrange neural flow tracing in conjunction with continuous-time neural evolution units. It generates flow field vectors by analyzing the tangent directions of geometric state points on the manifold surface, and uses these vectors as the slope basis to perform continuous-time integral derivation using a neural ordinary differential equation solver. This method essentially simulates the particle drift process in a physical field, overcoming the information loss and accumulated error problems caused by the discrete time step of traditional recurrent neural networks. It ensures that the derivation process strictly follows the continuous physical evolution law of the channel state, thereby achieving advanced and accurate prediction of time-varying channels in high-speed moving scenarios.

[0036] 3. This method introduces a real-time manifold deviation monitoring and geometric projection correction mechanism during the inference process, which can calculate the deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time. When a state jump that violates physical inertia occurs, the geometric projection correction term is used to force the state vector back to the manifold surface that conforms to physical laws, and this process does not require triggering the backpropagation of network weights. This mechanism effectively filters out non-physical abrupt changes caused by environmental noise or model approximation, ensuring the stability and robustness of the model in online operation, and solving the contradiction between high-precision prediction and low inference latency.

[0037] 4. This method maps the high-dimensional predicted state back to the physical space through a Riemann projection decoding unit, calculates the complex domain channel state information at future moments, and generates the optimal downlink precoding matrix accordingly. This mechanism pre-embeds compensation components for Doppler frequency shift and phase rotation at the transmitter, enabling the transmitted beam to achieve phase coherence superposition with the actual channel state at that time when it reaches the user equipment after experiencing propagation delay. This effectively solves the channel aging effect caused by rapid channel changes in high-speed railway or vehicle-to-everything (V2X) scenarios, and significantly improves the receiving signal-to-noise ratio and transmission reliability of the communication link. Attached Figure Description

[0038] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0041] Example 1:

[0042] Please see Figure 1 A method for transmitting and receiving signals in a wireless communication system based on an AI model, comprising:

[0043] The delayed coordinate embedding unit serves as the sensing entry point of the system. It receives historical pilot observation sequences from user equipment and uses the Takens embedding principle to map one-dimensional time-domain signals into geometric state points in high-dimensional phase space, thereby generating potential state vectors that characterize the topology of the channel dynamic system.

[0044] The potential state vector is input into the Lagrange neural flow tracing unit. The tangent direction of the geometric state point on the manifold surface is analyzed through the manifold derivative network. The flow field vector describing the velocity and direction of the channel state motion is calculated and passed to the next stage as the evolution control parameter.

[0045] The continuous-time neural evolution unit, as a solver of neural ordinary differential equations, receives the potential state vector as the initial position, performs continuous-time integral deduction based on the slope provided by the flow field vector, and calculates the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time during the deduction process. It uses geometric projection correction terms to force the projection of the state that violates physical inertia, thereby outputting the predicted potential state at a future specified time point.

[0046] The Riemann projection decoding unit uses a nonlinear mapping relationship to project the predicted latent state in the high-dimensional manifold space back into the Euclidean space and solve for the complex domain channel state information matrix at future time.

[0047] The closed-loop signal transmission control unit calculates the optimal downlink precoding matrix based on the complex domain channel state information matrix, and after mapping the service data stream to be transmitted to the optimal downlink precoding matrix, transmits the beam through the antenna array.

[0048] In this specific architectural configuration, the delay coordinate embedding unit is configured to perform a phase space reconstruction task. This unit receives historical pilot observation sequences from the user equipment, which are typically discrete time-series signals. Addressing the technical challenge that univariate observations cannot directly reflect the full picture of a high-dimensional nonlinear system, this unit executes the reconstruction logic based on Takens' embedding theorem: selecting a specific embedding dimension. and time delay Mapping a one-dimensional time-domain signal to A geometric state point in the Euclidean space; this geometric state point is defined as a potential state vector, which not only contains the current amplitude information of the signal, but also embeds the first and second derivatives of the signal and higher-order dynamic features, thereby constructing a numerical expression in phase space that can characterize the complete topological structure of the channel dynamic system.

[0049] Building upon this, the Lagrange neural flow tracking unit receives the latent state vector. The core function of this unit is to analyze the system's evolution rather than directly predicting numerical values. Its internally integrated manifold derivative network is trained to fit the tangent space of the phase space manifold; that is, after receiving the latent state vector representing the current position, it calculates the tangent vector of that point on the manifold surface. This calculation generates a flow field vector, a physical quantity describing the rate and direction of the channel state's evolution on the phase space manifold surface, characterizing the instantaneous evolution trend of the system at the current moment. This flow field vector is then passed to the next stage as an evolution control parameter, providing a deterministic derivative basis for subsequent continuous integration. ;

[0050] The continuous-time neural evolution unit is constructed as a solver of neural ordinary differential equations; the unit receives the latent state vector as initial values. Based on the slope field provided by the flow field vector, the system performs continuous time-axis evolution using an adaptive step-size numerical integration algorithm. During this dynamic evolution, the system executes a manifold constraint mechanism in parallel: it calculates the Euclidean distance between the current evolution trajectory point and the preset low-dimensional manifold surface in real time, i.e., the manifold deviation. When this deviation exceeds a confidence threshold set based on the statistical distribution of the training set, the system determines that the current state has experienced an abnormal disturbance that violates physical inertia and generates a negative gradient vector pointing towards the manifold surface as a geometric projection correction term. This correction term directly acts on the current integration state, forcibly projecting the deviated trajectory back onto the manifold surface, ensuring that the evolution path always conforms to the physical laws of the channel. This unit outputs a specified future time point. Predicting potential states;

[0051] The Riemann projection decoding unit performs the inverse mapping from the abstract space to the physical space; given that the predicted latent state resides in an abstract high-dimensional manifold coordinate system, this unit approximates the inverse mapping relationship using a pre-trained nonlinear decoding network; considering that rigorous mathematical bijection is difficult to achieve in lossy compression, this embodiment employs a strategy of minimizing the mean square error to train the decoder; the decoding network... The training objective is to minimize the physical channel parameters. The distance between the predicted recovery value and the loss function is:

[0052]

[0053] in, It is the Frobenius norm. The first term is the regularization weight coefficient, and the second term is the Jacobi regularization term, used to constrain the smoothness of the mapping to approximate the local properties of bijection.

[0054] Mapping high-dimensional state points back to physical Euclidean space, we can solve for the corresponding future moments. The complex domain channel state information matrix; based on this CSI matrix containing future fading characteristics, the closed-loop signal transmission control unit calculates the optimal downlink precoding matrix and performs beamforming operations to drive the antenna array to transmit an electromagnetic beam that matches the future channel state space;

[0055] Through the aforementioned technical approach based on phase space dynamics manifold reconstruction and continuous-time differential equation derivation, this embodiment achieves advanced and accurate prediction of time-varying channels in high-speed mobile communication scenarios. Compared with existing technologies based on recursive networks with discrete time steps, this scheme overcomes the information loss and accumulated errors caused by discrete sampling by solving continuous differential equations. By using manifold geometric constraint mechanisms to replace blind parameter fitting, the model's dependence on the number of parameters is reduced, and the contradiction between high-precision prediction and low inference latency is resolved, ensuring real-time response capability in extremely short coherence time in high-speed railway or V2X scenarios.

[0056] Example 2:

[0057] Using the Takens embedding principle, a one-dimensional time-domain signal is mapped to geometric state points in a high-dimensional phase space to generate a latent state vector, specifically including:

[0058] Construct a delay vector matrix to map historical pilot observations at different times into components on a high-dimensional coordinate axis, forming a numerical set that can reflect the whole system.

[0059] The numerical set is defined as a potential state vector and directly used as the initial position input to the Lagrange neural flow tracing unit to solve the problem that univariate observations cannot reflect the complete state of the channel dynamic system.

[0060] In the detailed processing flow of this embodiment, the delay coordinate embedding unit performs strict data spatialization operations; the system determines two key topology parameters: time delay. With Embedding Dimension Time delay The selection is usually determined based on the minimum point of mutual information, and the specific calculation formula is as follows: Calculate the observation sequence With delayed sequence Mutual information function Select The corresponding value when it first drops to a local minimum The value is used as the optimal delay time to ensure the independence between components; where, This represents the probability mass function obtained after performing histogram discretization and statistics on a continuous signal.

[0061] Embedding dimension The principle is based on the fact that the proportion of spurious nearest neighbors approaches zero, and different dimensions are calculated accordingly. The proportion of points whose distance changes abruptly within the neighborhood of the state vector. ;when When, the corresponding smallest integer That is, to confirm the embedding dimension This ensures that the phase space can be fully expanded without overlap; based on these two parameters, the system constructs a delay vector matrix, representing the current time step. Observations and historical moments Combination, mapped to one The coordinate vector in 3D space; this set of values ​​is defined as the latent state vector; this vector not only contains the current amplitude information, but also implicitly encodes the system's rate of change and acceleration information through historical delay components, thereby mathematically reconstructing the geometric structure that is differentially homeomorphic to the original dynamical system; this processing logic enables a single pilot observation to reflect the full picture of the system implied by Doppler frequency shift rate, relative motion of multipath scatterers, etc., providing a complete initial state input for the accurate derivative analysis of the subsequent Lagrange neural flow tracing unit.

[0062] Example 3:

[0063] The flow field vector is calculated by analyzing the tangent direction of the geometric state point on the manifold surface using a manifold derivative network, specifically including:

[0064] Configure the built-in manifold derivative network to fit the geometric tangent space, and receive the latent state vector as input;

[0065] Analyze the evolutionary physical inertia of the current state point on the surface of the phase space manifold, and generate control parameters that represent the evolutionary derivative of the system;

[0066] The control parameters are defined as flow field vectors, which are used to specify the motion trend of the channel state in phase space and serve as the core basis for the continuous-time neural evolution unit to perform integral operations.

[0067] In this embodiment, the operation of the Lagrange neural flow tracing unit is based on the principles of manifold geometry; its built-in manifold derivative network is constructed as a mapping function from phase space to tangent space. Specifically, this manifold derivative network adopts a residual fully connected network architecture; the network input layer dimension is... It contains 3 hidden layers, with each layer containing 100 neurons. The Swish activation function is used between hidden layers. To ensure the smoothness of the derivative, the network output layer does not use an activation function and directly outputs the flow field vector with the same dimension as the input. The training objective of the network is to minimize the one-step prediction error, and the loss function is defined as:

[0068]

[0069] The network receives the latent state vector as input, and after multiple nonlinear transformations, analyzes the local geometric characteristics of the current state point on the low-dimensional manifold surface. The system then analyzes the evolutionary physical inertia of this point in phase space, i.e., the inherent tendency of the system state to change over time, and generates a set of time derivatives representing the system's evolution. The control parameters; this set of parameters is defined as the flow field vector; this vector, in a physical sense, defines the instantaneous motion trend of the channel state in phase space, and its value is directly used as the differential equation in the continuous-time neural evolution unit. The right-hand item This forms the core mathematical basis for performing integral operations, ensuring that the subsequent deduction process strictly follows the physical evolution laws of the channel dynamics system.

[0070] Example 4:

[0071] Based on the slope provided by the flow field vector, a continuous-time integral derivation is performed, specifically including:

[0072] Numerical integration algorithms are used to simulate the particle drift process in physical fields;

[0073] The new position of the potential state vector after moving along the flow field trajectory under the action of the flow field vector is calculated, thereby realizing the continuous-time evolution simulation of the channel state.

[0074] During the simulation, the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface is calculated in real time. Geometric projection correction terms are used to force projection onto states that violate physical inertia, specifically including:

[0075] During the integral evolution process, the geometric distance between the current evolution trajectory position and the preset low-dimensional manifold surface is monitored in real time to obtain the manifold deviation value;

[0076] If the manifold deviation value exceeds the preset threshold, it is determined that a jump that violates physical inertia has occurred, and a geometric projection correction term is generated.

[0077] The geometric projection correction term is applied directly to the current integral state vector, forcing it back to the surface of the manifold that conforms to physical laws, and this correction process does not trigger the backpropagation of network weights.

[0078] In this preferred embodiment, the continuous-time neural evolution unit ensures high fidelity inference through the collaborative work of numerical analysis and geometric constraints. The system employs a high-order numerical integration algorithm, treating the input latent state vector as a point mass in phase space, and using the flow field vector as the instantaneous velocity field of this point mass to simulate its drift trajectory on the continuous time axis. This process calculates the state vector after the time increment. The new position reached later enables accurate simulation of the continuous evolution of the channel state;

[0079] To suppress non-physical abrupt changes caused by accumulated errors and environmental noise, the system synchronously performs manifold deviation monitoring in each integration step. The system utilizes a pre-trained manifold discriminator or distance function to calculate the Euclidean distance between the current evolution trajectory location and the preset low-dimensional manifold surface in real time, thus obtaining the manifold deviation value. The preset low-dimensional manifold surface M is implicitly defined by a pre-trained autoencoder; the bottleneck layer dimension of the autoencoder... The selection is based on the cumulative variance contribution rate of principal component analysis. The dimension value corresponding to a cumulative variance contribution rate of 95% in the training set data is selected as the dimensional value. The autoencoder constructs a training set using historical pilot observation data captured by a sliding window during communication. It is trained through unsupervised learning, with the training objective being to minimize the reconstruction error between the input and output. The loss function is... The self-encoder contains an encoder. and decoder manifold deviation value Defined as the current potential state vector The reconstruction error is calculated using the following formula:

[0080]

[0081] in, Represents the L2 norm. Represents the state vector reconstructed by the autoencoder; preset threshold Set as the 99th percentile value of the reconstruction error distribution of the training set samples;

[0082] The deviation value is compared with a preset threshold, which is a boundary value determined based on the statistical characteristics of the residual distribution of the training dataset on the manifold surface; if the manifold deviation value exceeds the threshold, the system determines that the current state has undergone a change that violates physical inertia; when At that time, the system freezes the autoencoder weights for the input state. Calculate the gradient of the reconstruction error The geometric projection correction term generated at this time Calculated according to the negative gradient direction of the manifold energy function:

[0083]

[0084] in, The projection step size, For the reconstruction point, The identity matrix corresponding to the dimension of the state vector. and These are the Jacobian matrices for the decoder and encoder, respectively; the corrected state update is... until Or reach the maximum number of iterations;

[0085] The system generates a geometric projection correction term based on the gradient descent direction. This correction term is a correction vector pointing to the manifold surface, which is directly superimposed on the current integral state vector, forcing it back to the manifold surface that conforms to physical laws. This correction process only involves adjusting the state vector values ​​during the deduction phase and does not trigger the backpropagation update of the neural network weights. Thus, while ensuring the stability of the model's online operation, it achieves zero-delay filtering of abnormal disturbances.

[0086] Example 5:

[0087] The Riemann projection decoding unit calculates the complex domain channel state information matrix for future time moments, specifically including:

[0088] Configure this unit as an output mapping layer to receive the predicted latent states in an abstract high-dimensional manifold space;

[0089] The predicted latent state is translated into physical communication parameters through a trained nonlinear network layer, and the complex domain channel state information matrix corresponding to time t+Δt is output.

[0090] This embodiment illustrates a mapping mechanism for regressing physical communication parameters from an abstract mathematical space. The Riemann projection decoding unit, as the system's output stage, is configured to perform a nonlinear manifold decoding task. Since the input predicted latent state is an abstract coordinate point located on a high-dimensional phase space manifold and does not directly correspond to a specific physical quantity, this unit utilizes a nonlinear network layer trained under end-to-end supervision to perform an inverse mapping operation. The nonlinear network layer is specifically configured as a fully connected multilayer perceptron structure, including an input layer, three hidden layers, and an output layer. The number of neurons in the hidden layers decreases sequentially, and the LeakyReLU activation function is used to handle nonlinear features. The output layer does not use an activation function to match the distribution in the real number domain. This operation accurately projects points on the high-dimensional manifold onto the physical parameter space, calculating the parameters corresponding to future times. The complex domain channel state information matrix contains detailed channel impulse responses or frequency responses of each subcarrier and antenna port at future times, providing complete physical layer parameter support for link adaptation of the communication system.

[0091] Example 6:

[0092] The closed-loop signal transmission control unit calculates the optimal downlink precoding matrix, specifically including:

[0093] Based on the predicted complex domain channel state information matrix for future moments, the calculation is performed using singular value decomposition or zero-forcing algorithm;

[0094] Generate the optimal downlink precoding matrix for the next time step to pre-compensate for Doppler shift and channel aging caused by high-speed movement.

[0095] After mapping the service data stream to be transmitted to the optimal downlink precoding matrix, the beam is transmitted through the antenna array, specifically including:

[0096] Execute formula Matrix operations, where, For the business data stream to be sent, For the optimal downlink precoding matrix, The weighted transmitted signal;

[0097] The system controls the transmission of weighted signals from a large-scale antenna array to form a beam pointing towards the user equipment.

[0098] In this embodiment, the closed-loop signal transmission control unit is responsible for converting the prediction gain into actual transmission performance improvement; the system receives the complex-domain CSI matrix at future time points. The system performs precoding calculations based on this matrix; the system uses the singular value decomposition algorithm to... The process involves decomposing the matrix and extracting the right singular vector as beamforming weights; or using a zero-forcing algorithm to calculate the pseudo-inverse matrix to eliminate inter-user interference. This calculation process generates the optimal downlink precoding matrix optimized for future channel characteristics. This matrix not only contains the weighting information for spatial reuse, but also implicitly contains the inverse compensation components for the phase rotation and amplitude fading caused by the Doppler frequency shift at future moments.

[0099] The system performs signal synthesis operations: matrix multiplication. ,in, This is the vector of the service data stream to be sent. For the above precoding matrix, This represents the transmit signal vector for each antenna port; the system drives the large-scale antenna array to transmit this weighted signal. Due to the signal The middle part is pre-configured to address transmission delay. The compensation information of the channel state after propagation can be coherently superimposed with the actual channel state at that time when the electromagnetic wave reaches the user equipment at high speed, forming a high-gain beam pointing to the user equipment. This mechanism effectively eliminates the channel aging effect under the traditional feedback mechanism and significantly improves the receiving signal-to-noise ratio and bit error rate performance in high-speed mobile scenarios.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for transmitting and receiving signals in a wireless communication system based on an AI model, characterized in that, include: The delayed coordinate embedding unit serves as the sensing entry point of the system. It receives historical pilot observation sequences from user equipment and uses the Takens embedding principle to map one-dimensional time-domain signals into geometric state points in high-dimensional phase space, thereby generating potential state vectors that characterize the topology of the channel dynamic system. The potential state vector is input into the Lagrange neural flow tracing unit. The tangent direction of the geometric state point on the manifold surface is analyzed by the manifold derivative network. The flow field vector describing the velocity and direction of the channel state motion is calculated and passed to the next level as the evolution control parameter. The continuous-time neural evolution unit, as a solver of neural ordinary differential equations, receives the potential state vector as the initial position, performs continuous-time integral deduction based on the slope provided by the flow field vector, and calculates the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time during the deduction process. It uses geometric projection correction terms to force the projection of the state that violates physical inertia, thereby outputting the predicted potential state at a future specified time point. The Riemann projection decoding unit uses a nonlinear mapping relationship to project the predicted latent state in the high-dimensional manifold space back into the Euclidean space and solves the complex domain channel state information matrix for future time moments. The closed-loop signal transmission control unit calculates the optimal downlink precoding matrix based on the complex domain channel state information matrix, and after mapping the service data stream to be transmitted with the optimal downlink precoding matrix, transmits the beam through the antenna array; By analyzing the tangent direction of the geometric state point on the manifold surface using a manifold derivative network, the flow field vector describing the velocity and direction of the channel state motion is calculated, specifically including: Configure the built-in manifold derivative network to fit the geometric tangent space, and receive the latent state vector as input; Analyze the evolutionary physical inertia of the current state point on the surface of the phase space manifold, and generate control parameters that represent the evolutionary derivative of the system; The control parameters are defined as flow field vectors, which are used to specify the motion trend of the channel state in phase space and serve as the core basis for the continuous-time neural evolution unit to perform integral operations. Continuous-time integration is performed based on the slope provided by the flow field vector, specifically including: Numerical integration algorithms are used to simulate the particle drift process in physical fields; The potential state vector is calculated to move along the flow field trajectory under the action of the flow field vector, thereby realizing the simulation of the evolution of the channel state in continuous time; The process of calculating the manifold deviation between the evolution trajectory and the preset low-dimensional manifold surface in real time during the deduction process, and using a geometric projection correction term to force projection of states that violate physical inertia, specifically includes: During the integral evolution process, the geometric distance between the current evolution trajectory position and the preset low-dimensional manifold surface is monitored in real time to obtain the manifold deviation value; If the manifold deviation value exceeds a preset threshold, it is determined that a jump that violates physical inertia has occurred, and a geometric projection correction term is generated. The geometric projection correction term is applied directly to the current integral state vector, forcing it back to the surface of the manifold that conforms to physical laws, and this correction process does not trigger the backpropagation of network weights.

2. The method according to claim 1, characterized in that, The process of mapping a one-dimensional time-domain signal to geometric state points in a high-dimensional phase space using the Takens embedding principle to generate a latent state vector specifically includes: Construct a delay vector matrix to map historical pilot observations at different times into components on a high-dimensional coordinate axis, forming a numerical set that can reflect the whole system. The numerical set is defined as a potential state vector and directly input as the initial position into the Lagrange neural flow tracing unit to solve the problem that univariate observations cannot reflect the complete state of the channel dynamics system.

3. The method according to claim 1, characterized in that, The Riemann projection decoding unit calculates the complex domain channel state information matrix for future time moments, specifically including: Configure this unit as an output mapping layer to receive the predicted latent states in an abstract high-dimensional manifold space; The predicted latent state is translated into physical communication parameters through a trained nonlinear network layer, and the complex domain channel state information matrix corresponding to time t+Δt is output.

4. The method according to claim 1, characterized in that, The closed-loop signal transmission control unit calculates the optimal downlink precoding matrix, specifically including: Based on the predicted complex domain channel state information matrix for future moments, the calculation is performed using singular value decomposition or zero-forcing algorithm; Generate the optimal downlink precoding matrix for the next time step to pre-compensate for Doppler shift and channel aging caused by high-speed movement.

5. The method according to claim 1, characterized in that, The step of mapping the service data stream to be transmitted to the optimal downlink precoding matrix and then transmitting the beam through the antenna array specifically includes: Execute formula Matrix operations, where For the business data stream to be sent, The optimal downlink precoding matrix is... The weighted transmitted signal; The large-scale antenna array is controlled to transmit the weighted transmission signal, forming a beam pointing towards the user equipment.

Images