CN-122001504-A - Channel time-frequency combined extrapolation method based on multi-layer perceptron hybrid architecture

CN122001504ACN 122001504 ACN122001504 ACN 122001504ACN-122001504-A

Abstract

The invention discloses a channel time-frequency joint extrapolation method based on a multi-layer perceptron hybrid architecture. Aiming at the problems that the acquisition cost of the channel data of the actual communication scene is high, the time consumption of the online training of the model is high, and the simulation data is difficult to cover the actual communication scene, the method utilizes a large-scale simulation channel data set to manufacture a pre-training model to extract the channel change characteristics, and then carries out online rapid adaptation according to the trace data of the actual measurement scene. Aiming at the characteristics of small delay spread and relatively small number of antennas of an indoor wireless channel, the invention applies the multi-layer perceptron hybrid architecture to the time-frequency domain extrapolation of a single-carrier broadband channel. The model can be quickly adapted to the actual communication scene by a small amount of data fine tuning. The method has the advantages of high adaptation speed, small data dependence, strong scene migration capability, high calculation parallelism and the like, and is suitable for developing edge deployment and rapidly adapting to actual scenes.

Inventors

ZHANG NIANZU
Zou Haofeng
GU JIAJING
WANG BAIWEN

Assignees

东南大学

Dates

Publication Date: 20260508
Application Date: 20260318

Claims (8)

1. A channel time-frequency joint extrapolation method based on a multi-layer perceptron hybrid architecture (MLP-Mixer) is characterized in that firstly, a large-scale channel data set is obtained by utilizing a public channel model simulation, so that the data can cover common indoor communication scenes to improve the generalization capability of a pre-training model as much as possible; Then, data preprocessing is carried out on the data set, an open-source simulation framework often ensures that the average power of single-shot channel data is 1, normalization processing is not needed to be carried out on the data under the condition, the multi-dimensional tensor of the single-carrier broadband channel data is converted into the two-dimensional tensor of the time sequence dimension and other characteristic dimensions from the time sequence dimension, the transceiver antenna dimension, the time delay tap dimension and the complex dimension, and meanwhile, the MLP-Mixer is favorable for separating stirring characteristics; then, constructing a channel time-frequency combined extrapolation model based on an MLP-Mixer architecture, carrying out time-position coding on the processed two-dimensional channel tensor to strengthen time sequence information of data, entering a time stirring layer and a characteristic stirring layer which are alternately stacked after coding, and respectively stirring the time dimension and other dimensions of the two-dimensional tensor; The method comprises the steps of carrying out pre-training by using measurement or simulation data of a channel, successfully obtaining a pre-training model with stronger generalization capability after training convergence, obtaining channel data of an actual communication scene by using a universal software radio peripheral, carrying out sliding time average preprocessing on the collected channel data to remove the influence of large-scale fading as far as possible, retaining the characteristic of small-scale fading, randomly extracting actual measurement channel data or extracting partial channel data at the initial moment of observation to correct the pre-training model, and realizing quick adaptation of the model and the scene.
2. The method for channel time-frequency joint extrapolation based on the multi-layer perceptron hybrid architecture of claim 1, comprising the steps of: Step 1, a large-scale channel data set is obtained by utilizing a public channel model simulation, namely, a large number of simulations are developed based on a standard channel model, a plurality of channel models, time delay expansion of a common indoor environment and indoor normal walking speed are covered, channel impulse responses are respectively sampled to different bandwidths to manufacture past, low-bandwidth-future and high-bandwidth single-carrier broadband channel data pairs; step 2, data remodeling, namely, the original channel data is remodeled into a two-dimensional tensor form, wherein the two-dimensional tensor form comprises a time dimension and other dimensions, and the other dimensions are mixed with information of a transmitting antenna, a receiving antenna, a time delay domain and a complex number; Step 3, simulation pre-training, namely inputting a training data set into a channel time-frequency extrapolation model based on a multi-layer perceptron hybrid architecture for training to obtain a pre-training model, namely performing time embedding on channel data, then alternately stirring the time dimension and other dimensions of the channel data, and performing average pooling in the original time dimension after full stirring to obtain a prediction result through a solution wharf; Step 4, pre-training weight loading, namely loading the weight of the pre-training model into a channel time-frequency combined extrapolation model to be adapted; And 5, performing online rapid adaptation, namely acquiring a small amount of channel observation data in an actual communication scene, performing online fine adjustment on a model to be adapted through preprocessing, and rapidly realizing the adaptation of the model to the actual communication scene.
3. The method is characterized in that the data set is generated through a standard channel model, covers various channel models aiming at indoor communication scenes, covers common indoor channel time delay expansion and moving speed, samples the common indoor channel time delay expansion and the moving speed to different bandwidths to obtain single-carrier broadband channel impulse responses with certain specifications, inputs data into the historical low-bandwidth MIMO channel impulse responses, outputs data into the future high-bandwidth MIMO channel impulse responses, collects actual measurement data, processes the acquired impulse responses under the corresponding bandwidths, and carries out sliding time averaging to ensure that the influence of large-scale fading is removed as much as possible and the characteristic of small-scale fading is kept.
4. The method for channel time-frequency joint extrapolation based on multi-layer perceptron hybrid architecture as recited in claim 3, wherein said data remodeling is performed by sequentially remodeling processed single-carrier wideband channel data into a sequence of time sequence, antenna index, delay tap and complex dimension Wherein, In order to make the number of time steps, For the pair of antenna pairs, Is the number of delay taps.
5. The method for channel time-frequency combined extrapolation based on multi-layer perceptron hybrid architecture as set forth in claim 4, wherein the channel time-frequency extrapolation model of the MLP-Mixer architecture first sends the historical channels into the time embedding layer for position coding, strengthens the sequence of channel observation data in the time dimension, and fully utilizes the time evolution characteristic of the channels.
6. The method for channel time-frequency combined extrapolation based on the multi-layer perceptron hybrid architecture as set forth in claim 5, wherein the time dimension and other dimensions of the alternately stirred channel data send the time-position encoded two-dimensional tensor to an alternately stacked time stirring layer and a feature stirring layer, and stir the time dimension and other dimensions of the two-dimensional tensor, respectively, wherein the time stirring layer is used for extracting a time evolution rule of the channel, and the feature stirring layer is used for relevant features of a space domain and a frequency domain of the channel.
7. The method for channel time-frequency joint extrapolation based on multi-layer perceptron hybrid architecture of claim 6, wherein channel data after being fully stirred by alternately stacked time stirring layers and feature stirring layers is averaged and pooled in a time dimension, so as to effectively reduce the number of parameters to avoid sinking into a locally optimal solution during training of a small amount of data.
8. The method for channel time-frequency combined extrapolation based on the multi-layer perceptron hybrid architecture as set forth in claim 7, wherein when the on-line fast adaptation of the model is performed, the model can be effectively modified by only moving the receiver to collect a small amount of channel data at a walking speed in an actual communication scene, and randomly extracting a small amount of data or extracting observation data in an initial period, so that the model is fast adapted to the actual communication scene.

Description

Channel time-frequency combined extrapolation method based on multi-layer perceptron hybrid architecture Technical Field The invention relates to the field of wireless channel prediction, in particular to a channel time-frequency joint extrapolation method based on a multi-layer perceptron hybrid (MLP-Mixer) architecture. Background In recent years, with the rapid development of sixth generation mobile communication technology (6G) and internet of things technology, a system has put near severe demands on the accuracy and timeliness of obtaining Channel State Information (CSI). The indoor environment presents rich multipath effect and complex space-time correlation characteristics due to complex building structure, dense scatterers and frequent personnel activities. In such highly dynamic and non-stationary environments, real-time prediction and extrapolation of wireless channel characteristics is expected to be a key technique for securing high quality communication links. The traditional channel analysis method is often based on the assumption of channel quasi-stability, and it is difficult to capture the nonlinear characteristic change of the channel caused by personnel walking or terminal movement in an indoor scene. The wireless channel is a high-dimensional random process in the time, frequency and spatial dimensions. The prior art often uses neural networks for analysis in dealing with such multidimensional coupling features. However, the current research method has some defects that a Convolutional Neural Network (CNN) is limited by a local receptive field, so that the frequency domain correlation under long time delay expansion and the time correlation of long time span are difficult to capture, and a cyclic neural network (RNN) is limited by calculation logic, so that the high-parallelism real-time reasoning is difficult to realize. The latest self-attention mechanism (transducer) computational complexity grows in square with increasing channel bandwidth and antenna dimensions, greatly limiting its deployment on computationally constrained edge terminals. For environments with obvious correlation with space-time in the indoor multipath effect, the scheme based on deep learning needs to be changed correspondingly. The indoor scene has obvious differentiation characteristics, and different room layouts and decoration materials can cause the statistical characteristics of channels to change greatly. Although the simulation data generated by using the standard channel model can provide physical prior for the model, the simulation data still has difficulty in covering rich actual communication scenes, and the model accuracy trained by simulation results in a real environment can be seriously reduced or even disabled. If the head training is performed by completely relying on measured data, the contradiction that a large amount of data is acquired, the model precision is high, but the data acquisition and training cost is high, the model convergence is slow, a small amount of data is acquired, the model precision is low, and the model is easy to fall into a local optimal solution can occur. Therefore, how to utilize the measured data with very small scale to realize the online rapid adaptation and high-precision extrapolation of the model to the specific scene under low computational complexity is a big difficulty faced by the current channel prediction and extrapolation. Disclosure of Invention The invention aims to provide a channel time-frequency combined extrapolation method based on a multi-layer perceptron hybrid architecture, which provides a method for predicting future high-bandwidth channel impulse response from historical low-bandwidth channel impulse response and provides a method for rapidly adapting to actual communication scenes. The invention has good performance on the time-frequency combined extrapolation of the indoor channel, can rapidly adapt to the actual communication scene only by a small amount of measured data, and has the advantages of high adaptation speed, small data dependence, strong scene migration capability, high calculation parallelism and the like. In order to achieve the aim of the invention, the channel time-frequency combined extrapolation method based on the multi-layer perceptron hybrid architecture provided by the invention firstly utilizes a public channel model to simulate and acquire a large-scale channel data set, and ensures that the data can cover common indoor communication scenes so as to improve the generalization capability of a pre-training model as much as possible; Then, data preprocessing is carried out on the data set, an open-source simulation framework often ensures that the average power of single-shot channel data is 1, normalization processing is not needed to be carried out on the data under the condition, the multi-dimensional tensor of the single-carrier broadband channel data is converted into the two-dimensional tensor of the time sequen