Orthogonal frequency division multiplexing (OFDM) channel estimation method based on Transform neural network

Abstract

The invention relates to an OFDM (Orthogonal Frequency Division Multiplexing) channel estimation method based on a Transform neural network, and belongs to the technical field of channel estimation methods. According to the method, the current channel state is estimated on the basis of a Transform attention mechanism by utilizing the correlation of the channel states of different sub-channels in the OFDM system among different frequencies in a frequency domain and the internal correlation of the channel states in a time domain. According to the technical scheme disclosed by the invention, the correlation of the channel state in two dimensions of frequency and time can be extracted at the same time, more accurate channel estimation is facilitated, and feature extraction in different scales is realized by using the convolutional neural network in the time domain, so that the processing capability of a long-time sequence can be enhanced with low complexity, and the channel estimation accuracy is improved. And a channel characteristic change rule is efficiently fitted, so that a channel environment with rapid change and dual selective fading can be more effectively coped with.

Description

Technical Field

[0001] This invention relates to an OFDM channel estimation method based on a Transformer neural network, belonging to the technical field of channel estimation methods. Background Technology

[0002] Orthogonal Frequency Division Multiplexing (OFDM) is an important multi-carrier transmission technology that plays a crucial role in modern wireless communication. Because OFDM effectively improves bandwidth utilization, exhibits good robustness against frequency-selective fading, and is easily integrated with other technologies, it is a key technology for 4G and 5G mobile communication systems.

[0003] In wireless communication systems, transmitted signals are affected by the characteristics of the propagation channel, resulting in distortion, fading, and path loss. Received signals are also susceptible to additive noise and interference, further degrading signal quality. To reliably recover transmitted signals in OFDM systems, accurate estimation of channel state information (CSI) is required for effective signal processing and data recovery.

[0004] Traditional channel estimation methods primarily rely on pilot information. This approach involves two steps: First, fixed pilot symbols, pre-agreed upon with the receiver, are inserted at the transmitter in a specific order. The receiver extracts the symbols at the pilot positions from the received signal and compares them with the original transmitted pilot symbols to obtain the Channel State Indicator (CSI) at those positions. Then, based on the CSI at the pilot positions, an interpolation algorithm is used to obtain the CSI at the data positions, ultimately yielding the CSI at all symbol positions. The structure of the inserted pilots is mainly divided into three categories: block pilots, comb pilots, and lattice pilots. The most common interpolation algorithms include Least Squares (LS) and Minimum Mean Square Error (MMSE). LS is simple to implement, but it doesn't consider noise during estimation, making it susceptible to interference and noise at low signal-to-noise ratios (SNR), limiting its applicability to a narrow range of SNRs. MMSE, on the other hand, utilizes second-order channel statistics and noise, resulting in better channel state estimation. However, it is more complex, requiring significant computational resources, and cannot be used when channel statistics are unavailable.

[0005] Deep learning technology has attracted attention from scholars in various fields due to its powerful feature extraction and pattern fitting capabilities. In using deep learning technology to solve the channel state estimation problem, the main solutions fall into two categories.

[0006] One type of approach analogizes the channel estimation problem to the super-resolution upsampling problem in image processing. These methods primarily utilize convolutional neural networks (CNNs) to construct an image super-resolution network and a denoising network. In the first stage, the channel estimation matrix, containing only pilot position channel state information, is treated as an image, and upsampling is performed using the super-resolution network. In the second stage, the denoising network filters out noise. This type of CNN-based approach has proven somewhat successful in channel estimation, but it does not utilize the long-term correlations of temporal frequency variations in the channel matrix. The second type of approach builds upon CNNs by further utilizing recurrent neural networks (RNNs) to capture temporal correlations. Systems based on long short-term memory (LSTM) recurrent neural networks have proven to outperform purely CNN-based systems. Methods based on gated recurrent units (GRUs), using computationally less complex RNN blocks, have also proven to outperform LSTM-based systems.

[0007] However, current RNN-based schemes utilize temporal correlations but do not consider how to simultaneously leverage the correlations between sub-channels in the frequency domain for more accurate estimations. Furthermore, as the time series lengthens, the processing power and complexity of RNNs become limited. Therefore, finding a way to simultaneously utilize temporal and frequency-domain correlations, and process long-term series with low complexity, to achieve more accurate channel estimations, presents a new challenge. Summary of the Invention

[0008] The purpose of this invention is to disclose an OFDM channel estimation method based on Transformer neural network. This method can simultaneously utilize the correlation between time domain and the correlation between frequency domain sub-channels, and process long time series with low complexity to achieve more accurate channel estimation.

[0009] To achieve the objectives of this invention, a method for OFDM channel estimation based on a Transformer neural network is provided, comprising the following steps: Step 1: Obtain the matrix containing only the channel state at the pilot position in the current OFDM time slot by inserting pilots. ,matrix Listed as Indicates the first Time within a time slot, behavior Represents different communication sub-channels in an OFDM communication system. for dimension; Step 2: Divide the time slot Previously obtained channel state matrix , , , After merging and splicing in chronological order, the trained convolutional neural network... After processing, the output matrix has the same number of rows as the channel state matrix. ; Step 3: Merge and stitch rows and the matrix obtained in step two Input to a trained Transformer-based neural network Output matrix Its dimensions and same; Step 4: Convert the matrix After transposing Input to a trained Transformer-based neural network Output matrix Its dimensions and same; Step 5: Convert the matrix sum matrix transpose matrix Simultaneously input a trained fully connected network Post-output matrix Its dimensions and same; Step Six: Utilize Channel state estimation data replacement at the middle pilot position The channel state estimation matrix is ​​obtained by taking the corresponding location data. ; Step 7: Repeat steps 1 to 7 to obtain the channel state matrix for each subsequent time slot.

[0010] Preferably, the convolutional neural network Neural networks based on Transformer Neural networks based on Transformer and fully connected neural networks These are all trained neural networks. The training process for these neural networks includes: Step 1: Construct a training dataset using known channel state data, through a sequence of known channel state matrices. The matrix is ​​obtained by removing the channel state values ​​at non-pilot locations. To obtain the training dataset , ,matrix Dimensions and structure same; Step 2: Randomly select a dataset of length from the training dataset. Continuous data sequences, utilizing and , , ,… Perform the operations according to steps one through five as described in claim 1, and the network output is obtained. Using loss function This represents the difference between the two, and updates the parameters of all neural networks based on the loss function until the loss function is reached. convergence; The loss function is the mean squared error. .

[0011] Preferably, in step two, a convolutional neural network is used. The processing procedure includes the following steps: Step 1: Obtain the channel state matrix before time slot N. , , ,… Merge and combine in chronological order; Step 2: For the merged matrix, divide the column vectors into equal parts. Groups A vector, each group is input sequentially. The number of rows in the output one-dimensional vector is the same as that in the channel state matrix. The outputs are merged into a matrix. ; Step 3: Obtain the results Repeat steps one and two again to obtain the matrix. until the matrix It contains only one column vector; Step Four: The output vector is obtained by taking the last set of time series matrices from the matrix; the others... Vectors on the same side of a matrix are concatenated row-wise to form a new matrix. , The number of column vectors contained therein is an integer multiple of the number of column vectors in the channel state matrix.

[0012] Preferably, the Transformer-based neural network Transformer-based neural networks A multi-head mechanism is adopted.

[0013] Preferably, the training dataset middle, .

[0014] Preferably, the number of vectors in each group when grouping column vectors is... .

[0015] Preferably, the The number of column vectors contained therein is an integer multiple of the number of column vectors in the channel state matrix.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. While utilizing time-domain sequence correlation, the correlation between frequency-domain sub-channels is also utilized, improving the estimation accuracy of channel state information.

[0017] This invention utilizes two Transformer-based neural networks simultaneously, wherein... Neural networks enable the extraction of channel state information (CSI) variation patterns from a time-domain perspective. The neural network extracts the variation pattern of CSI from the frequency domain perspective, and the two are connected through a fully connected network. The output then provides an estimate of the current CSI matrix. From a time-domain perspective, the CSI of an OFDM communication system is a time-series data point. From a frequency-domain perspective, there is also a strong correlation between sub-channels at different locations. A neural network architecture based on Transformer can consider the location information of sub-channels and increase the utilization of CSI correlations between frequency-domain sub-channels, thereby improving the accuracy of the current CSI estimation.

[0018] 2. Enhance the ability to process long-term series and improve the accuracy of state information estimation.

[0019] This invention employs a Transformer-based neural network. This method extracts the changing patterns of channel state information from a time-domain perspective. Compared to convolutional neural networks (RNNs), the Transformer-based neural network, employing an attention mechanism, can handle longer time sequences. The longer the CSI sequence processed, the better it helps to fit the changing patterns of CSI, thereby improving the estimation accuracy of CSI at the current moment.

[0020] 3. Effectively reduces the complexity of processing long-term series and minimizes the complexity increase brought about by improving estimation accuracy.

[0021] Before inputting the CSI sequence into the Transformer-based neural network architecture, this invention first processes it using a convolutional neural network. The processing flow is as follows: Figure 1 As shown, the CSI sequences are first grouped according to time sequence, and then input into a convolutional neural network respectively. Feature extraction is performed, and the data is reduced in dimensionality by the number of columns to obtain nodes with a coarser scale. Then, a convolutional neural network is used. The data is processed to obtain coarser-scale nodes. Nodes at different scales represent the characteristics of CSI information within different time ranges, thus forming a multi-resolution representation of the original CSI data. A matrix is ​​then constructed by selecting data closest to the current time at different scales. This data serves as input to the Transformer neural network. In this way, CSI data closer to the current time and with stronger correlation are represented at a small scale and high resolution, while CSI data farther away and with weaker correlation are represented at a large scale and low resolution. The combination of data at different scales forms a low-dimensional representation of the long-term series. CSI time-series data exhibits both long-term and short-term variation patterns, effectively capturing the correlation of long-term CSI series data, which is key to improving estimation accuracy. After passing through a convolutional neural network... Compared to directly using CSI time series information as input to the Transformer neural network, this method effectively reduces complexity. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the data processing flow of the convolutional neural network of this invention. Figure 2 The diagram shows the OFDM channel estimation method based on the Transformer neural network of this invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0024] Orthogonal Frequency Division Multiplexing (OFDM) technology is widely used, and accurate estimation of Channel State Information (CSI) is a prerequisite for effective signal processing and data recovery. Among traditional channel estimation algorithms, Least Squares (LS) is susceptible to interference and noise, and has a limited applicable signal-to-noise ratio range. Minimum Mean Square Error (MMSE) offers better channel state estimation, but it requires utilizing second-order channel statistics and noise, increasing complexity. Deep learning, as a new alternative, analogizes the channel estimation problem to super-resolution upsampling in image processing. It primarily utilizes Convolutional Neural Networks (CNNs) to construct image super-resolution and denoising networks. While this approach has yielded some results, it hasn't addressed the long-term correlation of CSI sequences over time and frequency. Introducing Recurrent Neural Networks (RNNs) to capture the temporal correlation of CSI sequences has achieved better results than using CNNs alone. However, challenges remain regarding how to capture the correlation between OFDM sub-channels in the frequency domain in parallel and how to process long-term sequences with low complexity to achieve accurate channel state estimation.

[0025] To address the problems existing in the prior art, in the embodiments disclosed in this invention, such as... Figure 2 As shown, the matrix containing only the channel state at the pilot position in the current OFDM time slot is first obtained by inserting pilot signals. The second is to utilize the trained convolutional neural network. Regarding what has already been obtained Pre-channel state estimation matrix , , , Process the data to output a matrix with the same number of rows as the channel state matrix. The third is merger. sum matrix Input to a trained Transformer-based neural network Output dimensions and Same matrix Fourth is to transform the matrix After transposing Input to a trained Transformer-based neural network Output dimensions and Same matrix The fifth is to make the matrix sum matrix transpose matrix Simultaneously input a trained fully connected network The output is the channel state matrix. ,use Channel state estimation data replacement at the middle pilot position The channel state estimation matrix is ​​obtained by taking the corresponding location data. .

[0026] Using convolutional neural networks The historical CSI sequence data is processed by first grouping the CSI sequences according to time sequence, with each group preferably consisting of CSI information from three consecutive time points, and then inputting them into a convolutional neural network. The output data is dimensionality-reduced by setting the number of columns to 1, resulting in larger-scale nodes. This is then processed using the same convolutional neural network to obtain even larger-scale nodes. Processing of the CSI sequence stops when only one node remains. Nodes at different scales represent the features of CSI information within different time ranges, thus forming a multi-resolution representation of the original CSI data. A matrix is ​​then constructed by selecting data closest to the current time at different scales. This is used as the input to the Transformer neural network. In the matrix... CSIs that are closer to the current time and have a stronger correlation are represented by small-scale, high-resolution data, while CSIs that are farther away from the current time and have a weaker correlation are represented by large-scale, low-resolution data. The combination of data at different scales forms a low-dimensional representation of long-term series. Compared with directly using CSI time-series information as input to the Transformer neural network, this can effectively reduce complexity while increasing the ability to process long-term series.

[0027] To address the correlations in the CSI time-domain sequences and the correlations between OFDM sub-channels in the frequency domain during parallel capture, two Transformer-based neural networks were employed. and . The input is a matrix It is used to capture the correlation within CSI time series; The input is a matrix The transpose of is used to capture the correlation between OFDM subchannels. and Parallel connections to a fully connected network The channel state matrix to be estimated is then output. Two neural networks and The number of heads is 4; fully connected neural network The number of layers is 3.

[0028] The technical solution disclosed in this invention requires convolutional neural networks. Transformer-based neural networks and Fully connected neural networks Conduct training.

[0029] During training, a training dataset is first constructed using a known sequence of channel state matrices. The matrix is ​​obtained by removing the channel state values ​​at non-pilot locations. To obtain the training dataset Training concentration .

[0030] During training, a dataset of length is randomly selected from the training dataset. Continuous data sequences, utilizing and , , ,… Based on the above, The second to fifth steps are performed sequentially, at which point the network output is obtained. Using loss function Update the parameters of all neural networks, repeat the operation, until the loss function is reached. Convergence. The loss function is... Mean squared error is a commonly used loss function with high efficiency, so it is used as the loss function.

[0031] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for OFDM channel estimation based on Transformer neural networks, characterized in that... The following is stated: Step 1: Obtain the matrix containing only the channel state at the pilot position in the current OFDM time slot by inserting pilots. ,matrix Listed as Indicates the first Time within a time slot, behavior Represents different communication sub-channels in an OFDM communication system. for dimension; Step 2: Divide the time slot Previously obtained channel state matrix , , , After merging and splicing in chronological order, the trained convolutional neural network... After processing, the output matrix has the same number of rows as the channel state matrix. ; Step 3: Merge and stitch rows and the matrix obtained in step two Input to a trained Transformer-based neural network Output matrix Its dimensions and same; Step 4: Convert the matrix After transposing Input to a trained Transformer-based neural network Output matrix Its dimensions and same; Step 5: Convert the matrix sum matrix transpose matrix Simultaneously input a trained fully connected network Post-output matrix Its dimensions and same; Step Six: Utilize Channel state estimation data replacement at the middle pilot position The channel state estimation matrix is ​​obtained by taking the corresponding location data. ; Step 7: Repeat steps 1 to 7 to obtain the channel state matrix for each subsequent time slot.

2. The OFDM channel estimation method based on Transformer neural network according to claim 1, characterized in that... The following is stated: The convolutional neural network Neural networks based on Transformer Neural networks based on Transformer and fully connected neural networks These are all trained neural networks. The training process for these neural networks includes: Step 1: Construct a training dataset using known channel state data, through a sequence of known channel state matrices. The matrix is ​​obtained by removing the channel state values ​​at non-pilot locations. To obtain the training dataset , ,matrix Dimensions and structure same; Step 2: Randomly select a dataset of length from the training dataset. Continuous data sequences, utilizing and , , ,… Perform the operations according to steps one through five as described in claim 1, and the network output is obtained. Using loss function This represents the difference between the two, and updates the parameters of all neural networks based on the loss function until the loss function is reached. convergence; The loss function is the mean squared error. .

3. The OFDM channel estimation method based on Transformer neural network according to claim 1, characterized in that... In step two, a convolutional neural network is used. The processing procedure includes the following steps: Step 1: Obtain the channel state matrix before time slot N. , , ,… Merge and combine in chronological order; Step 2: For the merged matrix, divide the column vectors into equal parts. Groups A vector, each group is input sequentially. The number of rows in the output one-dimensional vector is the same as that in the channel state matrix. The outputs are merged into a matrix. ; Step 3: Obtain the results Repeat steps one and two again to obtain the matrix. until the matrix It contains only one column vector; Step Four: The output vector is obtained by taking the last set of time series matrices from the matrix; the others... Vectors on the same side of a matrix are concatenated row-wise to form a new matrix. , The number of column vectors contained therein is an integer multiple of the number of column vectors in the channel state matrix.

4. The OFDM channel estimation method based on Transformer neural network according to claim 1, characterized in that... The above describes a neural network based on Transformer. Transformer-based neural networks A multi-head mechanism is adopted.

5. The OFDM channel estimation method based on Transformer neural network according to claim 2, characterized in that, Training dataset middle, .

6. The OFDM channel estimation method based on Transformer neural network according to claim 3, characterized in that, Group the column vectors and count the number of vectors in each group. .

7. The OFDM channel estimation method based on Transformer neural network according to claim 3, characterized in that, The number of column vectors contained therein is an integer multiple of the number of column vectors in the channel state matrix.

Images