CN-121690447-B - Method for transmitting and receiving signals of wireless communication system based on AI model

CN121690447BCN 121690447 BCN121690447 BCN 121690447BCN-121690447-B

Abstract

The invention relates to the technical field of wireless communication, in particular to a method for transmitting and receiving signals by a wireless communication system based on an AI model, which comprises delay coordinate embedding, nerve flow tracking, continuous evolution, riemann decoding and closed-loop control units, wherein the system utilizes Takens principle to map pilot frequency to high-dimensional phase space to reconstruct channel power topology, the core is that manifold derivative network is utilized to calculate flow field vectors, continuous time deduction is carried out through nerve ordinary differential equation, evolution deviation is eliminated by combining geometric projection correction terms, so that future channel state is accurately predicted, and CSI matrix is calculated based on prediction results to optimize beam emission.

Inventors

ZHANG WEIGANG
LIAN TIANFENG
XIA LIN
Hua Yongnan
LI NING
YANG JIANWEI
REN XIAOLI
GUO YONGPING

Assignees

天元瑞信通信技术股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260209

Claims (5)

1. A method for transmitting and receiving signals in a wireless communication system based on an AI model, comprising: The delay coordinate embedding unit is used as a perception entrance of the system, receives a historical pilot frequency observation sequence from user equipment, and maps one-dimensional time domain signals into geometrical state points in Gao Weixiang space by utilizing Takens embedding principle, so that potential state vectors representing the topological structure of the channel power system are generated; Inputting the potential state vector into a Lagrangian nerve flow tracking unit, analyzing the tangential direction of the geometric state point on a manifold curved surface through a manifold derivative network, calculating a flow field vector describing the movement speed and direction of the channel state, and transmitting the flow field vector to the next stage as an evolution control parameter; The continuous time nerve evolution unit is used as a neural ordinary differential equation solver, receives potential state vectors as initial positions, performs continuous time integral deduction according to the slope provided by the flow field vectors, calculates manifold deviation between an evolution track and a preset low-dimensional manifold surface in real time in the deduction process, and performs forced projection on a state violating physical inertia by using a geometric projection correction term so as to output a predicted potential state at a specified future time point; the Riemann projection decoding unit projects the predicted potential state in the high-dimensional manifold space back to the Euclidean space by utilizing a nonlinear mapping relation, and a complex domain channel state information matrix at a future moment is calculated; The closed loop signal transmission control unit calculates an optimal downlink precoding matrix based on the complex domain channel state information matrix, and transmits wave beams through the antenna array after mapping operation is carried out on the service data stream to be transmitted and the optimal downlink precoding matrix; analyzing tangential direction of the geometric state point on the manifold curved surface through a manifold derivative network, and calculating a flow field vector describing the movement speed and direction of the channel state, wherein the flow field vector specifically comprises the following steps: configuring a built-in manifold derivative network to fit a geometric tangent space, and receiving the potential state vector as an input; Analyzing the evolution physical inertia of the current state point on the phase space manifold surface, and generating control parameters representing the evolution derivative of the system; defining the control parameter as a flow field vector, wherein the flow field vector is used for prescribing the motion trend of a channel state in a phase space and is used as a core basis for executing integral operation by a continuous time neural evolution unit; continuous time integral deduction is carried out according to the slope provided by the flow field vector, and the method specifically comprises the following steps: Adopting a numerical integration algorithm to simulate a particle drift process in a physical field; calculating a new position of the potential state vector after moving along a flow field track under the action of the flow field vector, so as to realize evolution simulation of a channel state in continuous time; The method for calculating manifold bias of the evolution track and the preset low-dimensional manifold surface in real time in the deduction process comprises the following steps of: In the integral evolution process, monitoring the geometric distance between the current evolution track position and a preset low-dimensional manifold surface in real time to obtain a manifold deviation value; If the manifold deviation value exceeds a preset threshold, determining that jump violating physical inertia occurs, and generating a geometric projection correction term; And directly acting the geometric projection correction term on the current integral state vector, forcedly correcting the geometric projection correction term back to the manifold surface conforming to the physical rule, wherein the correction process does not trigger the back propagation of the network weight.
2. The method according to claim 1, wherein the mapping the one-dimensional time domain signal to geometrical state points in Gao Weixiang space using Takens embedding principle generates potential state vectors, specifically comprising: constructing a delay vector matrix, and mapping historical pilot observation values at different moments into components on a high-dimensional space coordinate axis to form a numerical value set capable of reflecting the overall view of the system; The numerical value set is defined as a potential state vector and is directly input into the Lagrangian nerve flow tracking unit as an initial position, so that the problem that the single variable observation cannot reflect the complete state of the channel power system is solved.
3. The method according to claim 1, wherein the Riemann projection decoding unit calculates a complex domain channel state information matrix at a future time, specifically comprising: configuring the unit as an output mapping layer, and receiving a predicted potential state in an abstract high-dimensional manifold space; and translating the predicted potential state into physical communication parameters through a trained nonlinear network layer, and outputting a complex domain channel state information matrix corresponding to the t+delta t moment.
4. The method according to claim 1, wherein the closed loop signaling control unit calculates an optimal downlink precoding matrix, specifically comprising: Calculating by using a singular value decomposition algorithm or a zero forcing algorithm based on the predicted complex domain channel state information matrix at the future time; and generating an optimal downlink precoding matrix at the next moment to pre-compensate Doppler frequency shift and channel aging caused by high-speed movement.
5. The method of claim 1, wherein after performing mapping operation on the service data stream to be sent and the optimal downlink precoding matrix, transmitting a beam through an antenna array specifically includes: Executing the formula Matrix operations of (1), wherein For the traffic data stream to be transmitted, For the optimal downlink precoding matrix, For the weighted transmit signal; and controlling the large-scale antenna array to transmit the weighted transmission signals to form a beam pointing to the user equipment.

Description

Method for transmitting and receiving signals of wireless communication system based on AI model Technical Field The invention relates to the technical field of wireless communication, in particular to a method for transmitting and receiving signals by a wireless communication system based on an AI model. Background Along with the development of wireless communication technology, high-speed mobile communication scenes such as high-speed railways, internet of vehicles and the like are increasingly popularized, nonlinear dynamics characteristics of a channel environment are increasingly complex, the complexity provides serious challenges for stability and reliability of signal transmission, particularly in terms of real-time acquisition and prediction of channel state information, a communication system generally relies on a scattered pilot frequency observation sequence to infer channel states and adopts a recursive neural network or linear interpolation method based on discrete time steps to conduct channel prediction and tracking at present, however, under the high-speed mobile environment, the existing method is limited by a discrete sampling mechanism, information loss and accumulated errors are inevitably caused, multipath scattering change caused by fast Doppler frequency shift is difficult to effectively cause channel aging effect very easily, in addition, a traditional model often depends on blind parameter fitting, the quantity of model parameters is large, the calculation efficiency is low, high-precision prediction and low-reasoning delay are difficult to be considered in extremely short coherent time, and the performance requirements of a real-time communication system cannot be met; Therefore, how to reconstruct the overall view of the channel power system by using limited historical observation data, overcome the limitation of discrete sampling, and realize the advanced accurate prediction and real-time compensation of the high-speed time-varying channel becomes a problem to be solved in the field. Disclosure of Invention In order to solve the above technical problems, the present invention provides a method for transmitting and receiving signals in a wireless communication system based on an AI model, and specifically, the technical scheme of the present invention includes: The delay coordinate embedding unit is used as a perception entrance of the system, receives a historical pilot frequency observation sequence from user equipment, and maps one-dimensional time domain signals into geometrical state points in Gao Weixiang space by utilizing Takens embedding principle, so that potential state vectors representing the topological structure of the channel power system are generated; Inputting the potential state vector into a Lagrangian nerve flow tracking unit, analyzing the tangential direction of the geometric state point on a manifold curved surface through a manifold derivative network, calculating a flow field vector describing the movement speed and direction of the channel state, and transmitting the flow field vector to the next stage as an evolution control parameter; The continuous time nerve evolution unit is used as a neural ordinary differential equation solver, receives potential state vectors as initial positions, performs continuous time integral deduction according to the slope provided by the flow field vectors, calculates manifold deviation between an evolution track and a preset low-dimensional manifold surface in real time in the deduction process, and performs forced projection on a state violating physical inertia by using a geometric projection correction term so as to output a predicted potential state at a specified future time point; the Riemann projection decoding unit projects the predicted potential state in the high-dimensional manifold space back to the Euclidean space by utilizing a nonlinear mapping relation, and a complex domain channel state information matrix at a future moment is calculated; the closed loop signal transmission control unit calculates an optimal downlink precoding matrix based on the complex domain channel state information matrix, and transmits wave beams through the antenna array after mapping operation is carried out on the service data stream to be transmitted and the optimal downlink precoding matrix. Preferably, the mapping the one-dimensional time domain signal to a geometrical state point in Gao Weixiang space by using Takens embedding principle to generate a potential state vector specifically includes: constructing a delay vector matrix, and mapping historical pilot observation values at different moments into components on a high-dimensional space coordinate axis to form a numerical value set capable of reflecting the overall view of the system; The numerical value set is defined as a potential state vector and is directly input into the Lagrangian nerve flow tracking unit as an initial position, so that the problem that the sing