CN-121984624-A - Channel parameter frequency domain extrapolation method for graph structure perception

CN121984624ACN 121984624 ACN121984624 ACN 121984624ACN-121984624-A

Abstract

The invention discloses a channel parameter frequency domain extrapolation method of graph structure perception, which comprises the steps of constructing a graph structure perception attention mechanism, constructing a channel parameter extrapolation network containing the graph structure perception attention mechanism, training the channel parameter extrapolation network by utilizing a channel parameter sequence data set containing known frequency points and corresponding target frequency points, inputting the channel parameter sequence of the known frequency points into the trained network, firstly extracting local features by a multi-scale convolution module, carrying out feature coding by an encoder integrating the graph structure perception attention mechanism, finally outputting a channel parameter prediction sequence of the target frequency points by a decoder, and effectively improving the prediction precision of the channel parameters in the frequency extrapolation process by joint modeling of multipath structure information. Experimental results show that under the condition of the same training data scale and network complexity, the method is superior to the existing method in the task of extrapolation of channel parameters.

Inventors

ZHANG NIANZU
GU JIAJING
Zou Haofeng
WANG BAIWEN
HUANG QINGYUE
PAN ZHIWEN

Assignees

东南大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (8)

1. A method for frequency domain extrapolation of channel parameters for graph structure perception, the method comprising the steps of: Step 1, constructing a diagram structure perception attention mechanism, wherein the mechanism constructs a frequency point-path association diagram based on the physical characteristics of channel multipath propagation, explicitly describes two kinds of association relations, namely the frequency domain correlation of the same path among different frequency points and the structure relation among different paths under the same frequency point, and fuses the structural information of the diagram as a learnable bias item into attention score calculation; Step 2, constructing a channel parameter extrapolation network containing the graph structure perception attention mechanism, wherein the network takes a multipath channel parameter sequence of a known frequency point as input and takes a multipath channel parameter sequence of an extrapolated target frequency point as output, and the network at least comprises a multi-scale feature extraction module for extracting multi-scale local features, a graph structure perception encoding module for deeply encoding structure association information and a decoding module for generating a target frequency point parameter sequence; Training the channel parameter extrapolation network by utilizing a channel parameter sequence data set containing known frequency points and corresponding target frequency points, wherein in the training process, a diagram structure perception attention mechanism and each module of the extrapolation network are simultaneously optimized, so that the network can learn a structure association priori between the frequency points and the multipath, and meanwhile, the whole network can effectively map the known frequency point sequence to the target frequency point sequence; and 4, inputting the channel parameter sequence of the known frequency point into a trained network, firstly extracting local characteristics by a multi-scale convolution module, then carrying out characteristic coding by an encoder integrated with a diagram structure perception attention mechanism, and finally outputting a channel parameter prediction sequence of the target frequency point by a decoder.
2. The method for frequency-domain extrapolation of channel parameters for graph structure awareness according to claim 1, wherein the graph structure awareness attention mechanism specifically comprises: According to the number of input frequency points Sum of multipath numbers Construction of the inclusion A frequency point-path association diagram of the node; the edge weights of the graph structure include: the longitudinal connection is that nodes of different frequency points of the same path are connected, and the rule of attenuation along with the increase of the frequency point interval is defined as: wherein Is the absolute value of the index difference of the frequency points, For the attenuation coefficient of F i , When (1) In the time-course of which the first and second contact surfaces, , The time is calculated according to the formula; transversely connecting nodes of different paths of the same frequency point, wherein the weight value is a fixed constant , 。
3. The method according to claim 2, wherein in step 1, the graph structure aware attention is implemented by the following formula: equation 1; Wherein: In order to query the matrix, In the form of a matrix of keys, In the form of a matrix of values, Is that And Is used for the vector dimension of (a), As a result of the scale factor being a leachable, An adjacent matrix of the frequency point-path association diagram is shown in the formula Representing a key matrix The attention mechanism employs a multi-headed attention structure to enhance feature modeling capabilities.
4. The method for extrapolating channel parameters in frequency domain for graphic structure perception according to claim 1, wherein the feature extraction module is a multi-scale feature extraction module, the module is a one-dimensional convolution structure and comprises at least three parallel convolution branches, each branch extracts local features through convolution kernels of different scales, and outputs of all branches are spliced through channel dimensions.
5. The method for extrapolating the channel parameter frequency domain perceived by the graph structure according to claim 1, wherein the graph structure perceived coding module performs joint coding on the input multi-frequency-point multipath channel characteristics based on a graph structure perceived attention mechanism, and introduces the structural information of the frequency point-path association graph into an attention weight calculation process in the coding process so as to constrain the characteristic interaction relationship, thereby enhancing the modeling capability of correlation between different frequency point channel parameters.
6. The method for frequency domain extrapolation of channel parameters based on structure perception feature output by encoding module as claimed in claim 1, wherein the decoding module generates the channel parameter sequence of the target frequency point step by step, and constrains the generated sequence information during generation to avoid introducing future information during prediction of channel parameters of the target frequency point, thereby realizing frequency domain extrapolation from channel parameters of known frequency point to channel parameters of the target frequency point.
7. The method according to claim 1, wherein the channel parameter extrapolation network further comprises an output mapping module for mapping the intermediate features output by the decoding module to channel parameters of the target frequency point, where the channel parameters include at least one or more of relative delay, relative power, and angle of arrival.
8. The method of extrapolation of channel parameters in frequency domain based on graph structure according to claim 1, wherein in step 4, the known frequency point is a continuous frequency point sequence, and the target frequency point is a continuous sequence of known frequency points The number of frequency points is one, And channel parameters of the input frequency point and the target frequency point are organized in a sequence form by taking a frequency point-path as a dimension, so that the integrity of multipath structure information is ensured.

Description

Channel parameter frequency domain extrapolation method for graph structure perception Technical Field The invention belongs to the field of communication and computers, and particularly relates to a channel parameter extrapolation method based on deep learning. Background In broadband wireless communication systems, channel parameters are used to characterize the impact of the wireless propagation environment on signal propagation characteristics, which are important bases for communication system design, performance assessment, and channel modeling. As communication systems are moving toward broadband and high frequency, channel parameters exhibit more complex variation characteristics between different frequency points. In engineering practice, channel parameters at different frequency points usually show a certain correlation, but are also commonly influenced by various factors such as propagation environment, multipath structure, frequency span, antenna configuration and the like. Limited by factors such as actual measurement cost, measurement time, and hardware conditions of measurement equipment, it is often difficult to obtain complete channel parameter data at all frequency points in practical application. The measurement or simulation can be performed only on a limited frequency point, so that extrapolation prediction is required to be performed on channel parameters of unmeasured frequency points according to channel parameters of known frequency points. The method has important engineering significance in application scenes such as wideband channel modeling, system level simulation, wireless resource allocation, new system communication system design and the like. Various solutions have been proposed for the problem of extrapolation of channel parameters. The conventional method is generally based on a physical propagation mechanism or a statistical channel model, and realizes estimation of channel parameters of unmeasured frequency points by constructing an analytical model or a statistical relation of channel parameters along with frequency changes. However, the method often depends on a specific propagation assumption or statistical distribution form, and model parameters are difficult to accurately acquire under the conditions of complex indoor environment, complex multipath structure or dynamic environment change, and the application range and prediction accuracy are limited to a certain extent. With the increase of computing power and the increase of the scale of measurement data, data-driven channel modeling and prediction methods are attracting attention. The deep learning method is used for describing complex channel characteristics by virtue of strong nonlinear modeling capability. Some studies use Convolutional Neural Networks (CNNs) to model channel parameters, and extract local relevant features of the channel parameters in the frequency dimension or parameter dimension through convolutional operations, and some studies introduce cyclic neural networks (RNNs) and their improved structures, such as long-short-term memory networks (LSTMs), to model sequence characteristics of the channel parameters in the frequency or time dimension. In addition, with the development of attention mechanisms, the transducer structure is introduced into channel modeling and prediction tasks due to its advantages in terms of sequence modeling and long-distance dependent modeling. The CNN is combined with the transducer structure in part of the research, the CNN is utilized to extract local characteristics, and then the transducer is utilized to model the cross-frequency point characteristics, so that the prediction or extrapolation of channel parameters is realized, and the flexibility and modeling capacity of the model are improved to a certain extent. However, existing methods based on CNN, LSTM or CNN-transducer usually treat multipath channel parameters as common sequences or feature vectors in the modeling process, and do not explicitly distinguish structural relationships between different multipaths, and do not fully characterize correlation characteristics of the same multipath between different frequency points. When the number of the multipath channels is large, the path structure is complex or the frequency point span is large, the method is difficult to effectively reflect the internal rule of the multipath structure along with the evolution of the frequency, and the accuracy of the extrapolation result is easily reduced or the stability is insufficient. Therefore, on the basis of the existing deep learning channel parameter extrapolation method, on the premise of not remarkably increasing the complexity of a model, an effective modeling mechanism for the association relation of the multipath structure is introduced, and the structural information of the multipath in the frequency dimension and the path dimension is fully utilized, so that the accuracy and the stability of channel parameter e