CN-122027060-A - Dynamic channel prediction method based on feature extraction and multi-scale attention

CN122027060ACN 122027060 ACN122027060 ACN 122027060ACN-122027060-A

Abstract

The invention discloses a dynamic channel prediction method based on feature extraction and multi-scale attention, which comprises the steps of collecting frequency domain impulse response data of a communication link in real time, preprocessing the frequency domain impulse response data and converting the frequency domain impulse response data into a real matrix of time domain impulse response, completing channel feature extraction and model training based on a modularized framework, wherein key channel feature characterization is strengthened through an SE module, performing multi-time-scale attention operation by utilizing an AMS module to accurately capture time domain features, finally generating a channel prediction model, deploying the trained prediction model to a system real-time processing unit, and realizing dynamic channel state prediction by combining real-time input data. The method optimizes the feature extraction efficiency while guaranteeing the prediction precision, remarkably improves the timeliness and suitability of channel prediction in complex communication scenes, enhances the perception capability of a low-altitude communication system to a dynamic channel environment, and has outstanding practicability and engineering application value.

Inventors

LI LIYAN
ZHAO RONG
Zhu keli

Assignees

浙江大学

Dates

Publication Date: 20260512
Application Date: 20260408

Claims (8)

1. The dynamic channel prediction method based on the feature extraction and the multi-scale attention is characterized by comprising the following steps: Constructing a channel data preprocessing module, acquiring original channel input data in a communication system, converting the original channel input data into a time domain impulse response form, constructing a real matrix, and finishing channel data preprocessing; Constructing a compression and excitation SE module, inputting the preprocessed real matrix into the SE module, and extracting the characteristics of the channel data and obtaining the reinforced channel characteristics through characteristic compression, self-adaptive weight distribution and characteristic recalibration; Constructing a self-adaptive multi-scale AMS module, setting a time scale parameter set of the AMS module, inputting channel characteristics output by the SE module into the AMS module, and extracting channel time domain dynamic characteristics under different time scales through a multi-scale attention mechanism; based on the multi-scale time domain dynamic characteristics output by the AMS module, a dynamic channel prediction result is obtained through calculation of a prediction output layer, and the prediction of channel state information is realized.
2. The method of claim 1, wherein the constructing a channel data preprocessing module, acquiring original channel input data in a communication system, converting the original channel input data into a time domain impulse response form, and constructing the original channel input data into a real matrix, and completing channel data preprocessing, comprises: acquiring Channel State Information (CSI) at the current moment, and parallelizing a transmitter and receiver antenna pair to construct a network input signal; converting the network input signal into a delay domain through Inverse Discrete Fourier Transform (IDFT) to obtain a delay domain signal; Normalizing the delay domain signal, mapping the delay domain signal into a real value tensor, and rearranging characteristic dimensions; And performing dimension expansion on the real value tensor through a full connection layer to obtain a real matrix adapting to the subsequent module.
3. The method of claim 2, wherein constructing a compression and excitation SE module, inputting the preprocessed real matrix into the SE module, and performing feature extraction on channel data and obtaining enhanced channel features through feature compression, adaptive weight distribution and feature recalibration processes, comprises: Setting compression coefficients, excitation function types and dimension parameters of a full connection layer of the SE module; inputting the preprocessed real matrix into a two-dimensional convolution layer, and obtaining a primary characteristic tensor through a ReLU activation function; Performing global average pooling operation on the preliminary feature tensor to realize feature compression and generate a feature vector with preset dimension; mapping the feature vectors through two full connection layers to simulate the correlation between time steps, and generating an attention weight tensor through a sigmoid function; And carrying out weighted scaling on the preliminary feature tensor by utilizing the attention weight tensor, and combining residual connection to obtain the reinforced channel feature.
4. The method of claim 3, wherein the constructing an adaptive multi-scale AMS module, setting a time scale parameter set of the AMS module, inputting channel characteristics output by the SE module into the AMS module, and extracting channel time domain dynamic characteristics at different time scales through a multi-scale attention mechanism, comprises: Defining a set containing M patch sizes, dividing an input feature sequence into patches with different time resolutions; Performing intra-patch attention calculation on each patch, and modeling local features through linear transformation and a cross attention mechanism to obtain intra-patch attention output; performing inter-patch attention calculation on the divided feature sequence, and modeling global correlation among different patches through a self-attention mechanism to obtain inter-patch attention output; And fusing the attention output in the patch with the attention output between the patches to obtain channel time domain dynamic characteristics under different time scales.
5. The method of claim 4, wherein the AMS module further comprises a multi-scale routing sub-module comprising: Decomposing an input sequence through a time decomposition sub-module to extract a periodic mode and a trend mode, wherein the periodic mode is obtained through Discrete Fourier Transform (DFT) and Inverse Discrete Fourier Transform (IDFT); Fusing the periodic mode, the trend mode and the original input sequence, and obtaining a time decomposition result through linear mapping; And generating path weights based on the time decomposition result introducing noise items, and executing top-k selection on the path weights to realize data self-adaptive patch size routing.
6. A dynamic channel prediction apparatus based on feature extraction and multi-scale attention, comprising: the first unit is used for constructing a channel data preprocessing module, acquiring original channel input data in a communication system, converting the original channel input data into a time domain impulse response form, constructing a real matrix and finishing channel data preprocessing; The second unit is used for constructing a compression and excitation SE module, inputting the preprocessed real matrix into the SE module, and extracting the characteristics of the channel data and obtaining the reinforced channel characteristics through characteristic compression, self-adaptive weight distribution and characteristic recalibration; the third unit is used for constructing a self-adaptive multi-scale AMS module, setting a time scale parameter set of the AMS module, inputting channel characteristics output by the SE module into the AMS module, and extracting channel time domain dynamic characteristics under different time scales through a multi-scale attention mechanism; And the fourth unit is used for calculating a dynamic channel prediction result through a prediction output layer based on the multi-scale time domain dynamic characteristics output by the AMS module so as to realize the prediction of channel state information.
7. An electronic device comprising a processor and a memory; wherein the processor runs a program corresponding to executable program code stored in the memory by reading the executable program code for implementing the method according to any one of claims 1-5.
8. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the program, when executed by a processor, implements the method according to any one of claims 1-5.

Description

Dynamic channel prediction method based on feature extraction and multi-scale attention Technical Field The invention relates to the technical field of wireless communication, in particular to a dynamic channel prediction method based on feature extraction and multi-scale attention. Background With the large-scale application of the fifth generation (5G) mobile communication technology and the prospective research and development of the sixth generation (6G) technology, emerging services such as high-definition video, internet of vehicles, industrial internet and the like continue to emerge, and more stringent requirements are put on performance indexes such as transmission rate, time delay, reliability and the like of a wireless communication system. The channel is used as a core carrier for wireless signal transmission, the state of the channel directly determines the quality of a communication link, the wireless channel has time-varying, space-varying and frequency-varying dynamic characteristics, and under high-frequency communication scenes such as low-altitude communication, the channel fading is severe, the coherence time is short, and the state instability is further aggravated, so that accurate channel state prediction becomes a key premise for improving the system performance. The channel prediction technology can infer the channel state at the future moment by analyzing and mining historical channel data, and can provide support for optimizing transmission parameters in advance and resisting negative effects of dynamic changes, so that the transmission reliability and the frequency spectrum efficiency are improved. The traditional statistical model is built based on priori statistical characteristics, has low calculation complexity, is poor in adaptability to complex and changeable scenes, is difficult to meet scene requirements of high-frequency communication, multi-user concentration and the like, and the data driving method relies on characteristic learning and nonlinear fitting capabilities of a deep learning technology, so that the traditional channel prediction method becomes a field research hotspot, and is mainly implemented by adopting models such as a convolutional neural network, a cyclic neural network and the like. However, the existing data driving scheme still has obvious defects that firstly, the key feature extraction capability is limited, noise and redundant information are easy to interfere, the feature characterization effectiveness is insufficient, secondly, the multi-time scale feature components in the dynamic change of a channel are difficult to adapt, the change rule of different time dimensions cannot be comprehensively captured by a single scale extraction mechanism, and the prediction precision and the generalization performance are restricted. Especially in a high-speed dynamic scene, the channel change rate is high, the influence factors are complex, the defects are further amplified, and the channel prediction with high precision and low time delay is difficult to realize. Disclosure of Invention The invention mainly aims to provide a dynamic channel prediction method based on feature extraction and multi-scale attention. Aiming at the problems of strong time-varying characteristics and complex multipath effect of a channel in a MIMO system, the dynamic channel prediction method based on feature extraction and multi-scale attention is provided, so that accurate and efficient prediction of channel states is realized, and the transmission reliability of a communication link is improved. It is another object of the present invention to provide a dynamic channel prediction apparatus based on feature extraction and multi-scale attention. A third object of the present invention is to propose an electronic device. A fourth object of the present invention is to propose a non-transitory computer readable storage medium. To achieve the above object, an embodiment of a first aspect of the present invention provides a dynamic channel prediction method based on feature extraction and multi-scale attention, including: Constructing a channel data preprocessing module, acquiring original channel input data in a communication system, converting the original channel input data into a time domain impulse response form, constructing a real matrix, and finishing channel data preprocessing; Constructing a compression and excitation SE module, inputting the preprocessed real matrix into the SE module, and extracting the characteristics of the channel data and obtaining the reinforced channel characteristics through characteristic compression, self-adaptive weight distribution and characteristic recalibration; Constructing a self-adaptive multi-scale AMS module, setting a time scale parameter set of the AMS module, inputting channel characteristics output by the SE module into the AMS module, and extracting channel time domain dynamic characteristics under different time scales through a