CN-121665278-B - Terminal-to-terminal wireless communication system and method based on AI

CN121665278BCN 121665278 BCN121665278 BCN 121665278BCN-121665278-B

Abstract

The invention relates to the technical field of wireless communication, in particular to an end-to-end wireless communication system and method based on AI, which are used for collecting signals, channels and hardware data in a real scene to construct an original data set, constructing a parameter-adjustable joint distortion characterization by a hybrid modeling method of fusion of differentiable physical simulation and data-driven residual error learning based on the data set, combining the characterization with the internal state of a communication process, dynamically adjusting parameters by using reinforcement learning ideas to generate an antagonism distortion mode, carrying out multi-round antagonism training on the communication process by using the mode, alternately optimizing communication parameters and distortion generation strategies until reaching a stable robust state, and finally deploying the trained communication process on an actual link, realizing parameter self-adaption matching and on-line fine adjustment by environment perception.

Inventors

ZHANG WEIGANG
FAN YI
DUAN HONGTAO
XIA LIN
LI FEIZHOU
HU JINGBO
WANG HUAN
KANG GANG

Assignees

天元瑞信通信技术股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260206

Claims (8)

1. An AI-based end-to-end wireless communication method, comprising the steps of: S1, acquiring signal sequences transmitted between a transmitting end and a receiving end in different actual scenes, and recording corresponding channel state information and hardware working state data to form an original data set; s2, based on an original data set, constructing an integrated joint distortion characterization with continuously adjustable parameters by a hybrid modeling method of fusing differentiable physical simulation and data-driven residual error learning, wherein the method specifically comprises the following steps: Constructing a differentiable physical simulation structure based on an electromagnetic propagation equation, wherein the differentiable physical simulation structure takes channel parameters and hardware parameters extracted from original data set in the previous link as input and outputs basic distortion response; training a residual error learning network by using the original data set, wherein the residual error learning network takes a basic distortion response and a corresponding actual measurement signal of a previous link as inputs to learn and output high-dimensional residual error characteristics; weighting and fusing the basic distortion response and the high-dimensional residual error characteristic to form a joint distortion characterization with continuously adjustable parameters; s3, taking the joint distortion characterization and the internal state information of the current end-to-end communication process as input, and adjusting a plurality of parameters in the joint distortion characterization in real time to generate an antagonistic distortion mode for reducing the performance of the current communication process; S4, performing multi-round training on the end-to-end communication process by utilizing the opposite distortion mode, optimizing parameters of the communication process by using the current distortion mode in each round, and updating a generation strategy of the distortion mode based on performance feedback of the communication process after optimization until the communication process reaches a stable state under various distortion conditions; S5, applying the end-to-end communication process obtained after training to an actual wireless communication link to complete the whole intelligent transmission and recovery from the sending signal to the receiving signal, wherein the method specifically comprises the following steps: monitoring the real-time environment state of an actual wireless communication link, and extracting the physical channel characteristics and the hardware working point of the current link; according to the instant environment state, self-adaptively selecting or fusing a parameter combination matched with the current condition from various parameter configurations accumulated in the communication process in the training process; in an actual wireless communication link, using the end-to-end communication process initialized by the selected parameter combination to perform signal transmission, and collecting transmission effect data of the signal transmission in real time; Based on the transmission effect data, the parameter combination is adjusted online until the performance index of the whole intelligent transmission and recovery reaches the preset optimal working threshold.
2. The AI-based end-to-end wireless communication method of claim 1, wherein the training the residual learning network using the raw data set specifically comprises: Carrying out alignment processing on the original data set, and accurately matching the basic distortion response with the corresponding actual measurement signal according to the time stamp and the signal frame structure to form paired training samples; constructing a residual error learning network with a bidirectional interaction structure, wherein the residual error learning network respectively processes basic distortion response and actual measurement signals through parallel characteristic extraction channels, and fuses two paths of characteristics in a deep layer through a cross attention mechanism so as to calculate high-dimensional residual error characteristics; and adopting a progressive training strategy based on course learning, performing preliminary training on a residual error learning network by using training samples with weaker distortion degrees, and gradually introducing training samples with stronger distortion degrees until the network can stably output high-dimensional residual error characteristics covering all distortion ranges.
3. The AI-based end-to-end wireless communication method of claim 1, wherein S3 specifically comprises: extracting internal state information representing instantaneous decoding confidence and signal characteristic distribution in the current end-to-end communication process; Calculating a performance degradation gradient of the communication process under the current joint distortion characterization based on the internal state information; according to the performance degradation gradient, channel multipath time delay expansion, hardware nonlinear strength and noise base parameters in the joint distortion characterization are synchronously and correlatively adjusted; Substituting the adjusted parameter set into the joint distortion characterization, synthesizing in real time and outputting an antagonistic distortion mode which locally maximizes the communication error rate.
4. The AI-based end-to-end wireless communication method of claim 3, wherein calculating a performance degradation gradient of the communication process under a current joint distortion characterization comprises: extracting confidence distribution of the current decoding process and characteristic statistics of the received signals from the internal state information; Constructing a scalar evaluation function reflecting the instantaneous robustness level of the communication process based on the confidence distribution and the feature statistics; Applying deterministic disturbance to key parameters in the joint distortion characterization, and obtaining corresponding variable quantity of a scalar evaluation function through forward calculation; And calculating to obtain the performance degradation gradient according to the ratio relation of the variable quantity and the deterministic disturbance.
5. The AI-based end-to-end wireless communication method of claim 1, wherein S4 specifically comprises: Performing a competitive adaptive modulation training for the peer-to-peer communication process based on the current antagonistic distortion pattern; Collecting performance feedback of the trained communication process in an antagonistic distortion mode, simultaneously evaluating the challenge strength of the antagonistic distortion mode to the communication process, and updating strategy parameters for generating the antagonistic distortion mode according to the bidirectional evaluation result; repeating the training and updating steps, monitoring the dynamic balance relation between the performance feedback and the challenge intensity in real time, and judging that the communication process reaches a stable state when both are maintained in a preset stable interval in continuous multi-round training.
6. The AI-based end-to-end wireless communication method of claim 5, wherein updating policy parameters for generating an antagonistic distortion pattern comprises: Normalizing the performance feedback and the challenge intensity to obtain quantized performance degradation indexes and distortion challenge levels respectively; Inputting the performance degradation index and the distortion challenge level into a dynamic strategy updating function, and calculating a strategy adjustment vector by the dynamic strategy updating function; And using the strategy adjustment vector to perform one iteration update on the strategy parameters according to which the distortion mode is generated.
7. The AI-based end-to-end wireless communication method of claim 1, wherein the online adjustment of the parameter combination specifically comprises: Analyzing the transmission effect data in real time, and extracting the error rate trend of the current signal transmission and the distortion characteristics of the received signal constellation diagram; matching the error rate trend and the distortion characteristic with a lightweight experience library constructed by historical successful adjustment records to obtain a group of alternative parameter fine adjustment vectors; Selecting one with the minimum evaluation cost from the candidate parameter fine tuning vectors, and executing superposition operation on the parameter combination to generate new online working parameters; And continuing to transmit by using the new online working parameters until the calculated efficiency index is continuously iterated for a plurality of times without lifting, and confirming that the optimal working threshold is reached.
8. An AI-based end-to-end wireless communication system for performing the AI-based end-to-end wireless communication method of any of claims 1-7, comprising: the data processing module is used for acquiring signal sequences transmitted between the transmitting end and the receiving end in different actual scenes and recording corresponding channel state information and hardware working state data to form an original data set; the joint distortion characterization construction module is used for constructing an integrated joint distortion characterization with continuously adjustable parameters by a hybrid modeling method of fusing differentiable physical simulation and data-driven residual error learning based on an original data set, and specifically comprises the following steps: Constructing a differentiable physical simulation structure based on an electromagnetic propagation equation, wherein the differentiable physical simulation structure takes channel parameters and hardware parameters extracted from original data set in the previous link as input and outputs basic distortion response; training a residual error learning network by using the original data set, wherein the residual error learning network takes a basic distortion response and a corresponding actual measurement signal of a previous link as inputs to learn and output high-dimensional residual error characteristics; weighting and fusing the basic distortion response and the high-dimensional residual error characteristic to form a joint distortion characterization with continuously adjustable parameters; the antagonistic distortion mode generation module takes the internal state information of the joint distortion characterization and the current end-to-end communication process as input, and adjusts multiple parameters in the joint distortion characterization in real time to generate an antagonistic distortion mode for reducing the performance of the current communication process; The opposite iteration training module performs multiple rounds of training on the end-to-end communication process by utilizing the opposite distortion mode, optimizes parameters of the communication process by the current distortion mode in each round, and updates a generation strategy of the distortion mode based on performance feedback of the communication process after optimization until the communication process reaches a stable state under various distortion conditions; The intelligent communication deployment module is used for applying the end-to-end communication process obtained after training to an actual wireless communication link to complete the whole intelligent transmission and recovery from a signal sending process to a signal receiving process, and specifically comprises the following steps: monitoring the real-time environment state of an actual wireless communication link, and extracting the physical channel characteristics and the hardware working point of the current link; according to the instant environment state, self-adaptively selecting or fusing a parameter combination matched with the current condition from various parameter configurations accumulated in the communication process in the training process; in an actual wireless communication link, using the end-to-end communication process initialized by the selected parameter combination to perform signal transmission, and collecting transmission effect data of the signal transmission in real time; Based on the transmission effect data, the parameter combination is adjusted online until the performance index of the whole intelligent transmission and recovery reaches the preset optimal working threshold.

Description

Terminal-to-terminal wireless communication system and method based on AI Technical Field The invention relates to the technical field of wireless communication, in particular to an AI-based end-to-end wireless communication system and method. Background As wireless communication technology moves toward higher frequency band, more complex scenarios, traditional modular communication system designs face significant challenges. The conventional method generally carries out discrete design and optimization on links such as a transmitter, channel coding and decoding, modulation and demodulation and the like based on a definite mathematical model (such as shannon theory and a specific modulation and coding scheme). This mode performs well in ideal or steady state channels, but is difficult to effectively cope with complex, time-varying and difficult to accurately model "joint distortions" in dense urban environments, high-speed movement, etc., caused by multipath, doppler effects and radio frequency hardware inherent impairments (e.g., power amplifier nonlinearities, phase noise) coupling. In recent years, artificial Intelligence (AI) technology has provided a new paradigm for communication physical layer design, namely, end-to-end communication systems based on deep neural networks. The system regards the transmitter and the receiver as a whole for joint optimization, and aims to automatically find an efficient and reliable coding and decoding strategy by learning rules in data and bypassing an explicit mathematical model. The traditional method cannot accurately model and reproduce the fast time-varying channel characteristics caused by high-speed movement, complex reflection and the like in a real wireless environment, and complex joint distortion generated by real-time coupling action between the fast time-varying channel characteristics and the inherent damage of hardware such as nonlinearity of a power amplifier, phase noise and the like, so that once the system trained in an ideal simulation environment is deployed in a real scene (such as a dense urban Internet of vehicles), failure risks such as decoding failure, bit error rate rising and the like appear when the system is faced with unknown distortion combination. Disclosure of Invention The present invention is directed to an AI-based end-to-end wireless communication system and method for solving the above-mentioned problems. The aim of the invention can be achieved by the following technical scheme: An AI-based end-to-end wireless communication method, comprising the steps of: S1, acquiring signal sequences transmitted between a transmitting end and a receiving end in different actual scenes, and recording corresponding channel state information and hardware working state data to form an original data set; s2, based on an original data set, constructing an integrated joint distortion characterization with continuously adjustable parameters by a hybrid modeling method of fusing differentiable physical simulation and data-driven residual error learning; s3, taking the joint distortion characterization and the internal state information of the current end-to-end communication process as input, and adjusting a plurality of parameters in the joint distortion characterization in real time to generate an antagonistic distortion mode for reducing the performance of the current communication process; S4, performing multi-round training on the end-to-end communication process by utilizing the opposite distortion mode, optimizing parameters of the communication process by using the current distortion mode in each round, and updating a generation strategy of the distortion mode based on performance feedback of the communication process after optimization until the communication process reaches a stable state under various distortion conditions; And S5, applying the end-to-end communication process obtained after training to an actual wireless communication link to complete the whole intelligent transmission and recovery from the signal transmission to the signal reception. The invention further provides a further scheme that the S2 specifically comprises the following steps: Constructing a differentiable physical simulation structure based on an electromagnetic propagation equation, wherein the differentiable physical simulation structure takes channel parameters and hardware parameters extracted from original data set in the previous link as input and outputs basic distortion response; training a residual error learning network by using the original data set, wherein the residual error learning network takes a basic distortion response and a corresponding actual measurement signal of a previous link as inputs to learn and output high-dimensional residual error characteristics; and carrying out weighted fusion on the basic distortion response and the high-dimensional residual error characteristic to form a joint distortion characterization with continuously adj