阅读说明：本技术 基于深度学习的ofdm信道估计方法 (OFDM channel estimation method based on deep learning ) 是由 高明 廖覃明 李靖 潘毅恒 黄凤杰 于 2020-03-12 设计创作，主要内容包括：本发明公开了一种基于深度学习的OFDM信道估计方法,主要解决现有技术信道估计质量差或实现复杂度太高的问题。其方案为：在接收端,获取时域信号y并进行预处理,得到接收信号导频位置的频域信号Y<Sub>P</Sub>；利用全连接层神经网络搭建信道估计模型CE-Net,并对其训练；利用现实环境的数据进行迁移训练；将CE-Net置于接收端,用于线上的信道估计。本发明降低了信道估计的实现复杂度,显著地提高了信道估计质量,可用于梳状导频模式下的OFDM通信系统。(The invention discloses an OFDM channel estimation method based on deep learning, which mainly solves the problems of poor channel estimation quality or too high implementation complexity in the prior art. The scheme is as follows: at a receiving end, acquiring a time domain signal Y and preprocessing the time domain signal Y to obtain a frequency domain signal Y of a pilot frequency position of a received signal P (ii) a Utilizing a full-connection layer neural network to build a channel estimation model CE-Net, and training the CE-Net; carrying out migration training by using data of a real environment; CE-Net is placed at the receiving end for on-line channel estimation. The invention reduces the complexity of channel estimation, obviously improves the channel estimation quality, and can be used for OFDM communication system in comb pilot mode.)

1. An OFDM channel estimation method based on deep learning is characterized by comprising the following steps:

(1) receiving the time domain signal Y by a receiving end and preprocessing the time domain signal Y to obtain a frequency domain signal Y of a pilot frequency position of the received signal_P；

(2) And (3) building a channel estimation model CE-Net by utilizing a full connection layer neural network:

(2a) the channel estimation model CE-Net comprises an input layer, a plurality of hidden layers and an output layer;

(2b) frequency domain signal Y of pilot frequency position of received signal_PThe output of the model is an estimated vector of the true channel vector H as input to the model

(3) Training a channel estimation model CE-Net by using sample data to obtain model parameters:

(3a) generating a time domain interface by using a Matlab software simulation platformAn input sample set formed by the received signal Y and a label sample set formed by a corresponding real channel vector H are preprocessed according to the mode described in the step (1) to obtain a frequency domain signal Y of a pilot frequency position of the received signal_PAnd the sample data composed of the corresponding real channel vector H;

(3b) vector estimation of channelTaking a mean square error function between the real channel vector H and the channel estimation model CE-Net as a cost function of the channel estimation model CE-Net, and performing offline training on the model by using sample data to minimize the cost function to obtain model parameters, namely weight W and bias b; w, b is input into the model to obtain a trained channel estimation model CE-Net;

(4) taking a sample set consisting of a frequency domain signal of a pilot frequency position of a received signal selected from a real environment and a corresponding channel vector as target domain data, carrying out migration training on a trained channel estimation model CE-Net, and adjusting parameters of the channel estimation model CE-Net;

(5) estimating an on-line channel by using the channel estimation model CE-Net obtained in the step (4):

(5a) placing the channel estimation model obtained in the step (4) at a receiving end;

(5b) in the on-line test or use process, the receiving end obtains the frequency domain signal Y of a certain receiving signal pilot frequency position according to the mode described in the step (1)_PInput to the model to obtain a channel estimation vector

2. The method of claim 1, wherein (1) the receiving end receives and pre-processes the time domain signal Y to obtain a frequency domain signal Y of the pilot position of the received signal_PIt is implemented as follows:

(1a) the receiving end obtains a time domain receiving signal y, and removes a cyclic prefix CP and a discrete prefix CP from the y in sequenceFourier transform DFT to obtain frequency domain receiving signalWherein C represents a set of complex numbers, N_cThe frequency domain received signal Y contains information Y of pilot frequency position for the number of sub-carriers_PAnd information Y of data position_D；

(1b) Let the number of pilots be N_P＝N_c/D_PIn which N is_cIs the number of subcarriers, D_PIs a pilot interval; because the pilot frequency position and the pilot frequency value are known in advance by the sending end and the receiving end, and the pilot frequency values are all complex v, the frequency domain receiving signal of the pilot frequency position is extracted from the frequency domain receiving signal Y according to the known pilot frequency position

3. The method of claim 1, wherein the structure of the channel estimation model CE-Net in (2a) is as follows:

the channel estimation model CE-Net comprises an input layer, a plurality of hidden layers and an output layer;

the input layer and the first hidden layer are both 2N_PA neuron of which N is_PIs the number of pilot frequencies; the output layer is composed of 2N_cA neuron of which N is_cIs the number of subcarriers.

4. The method of claim 3, wherein the activation function of the first hidden layer adopts a Sigmoid function, and the activation functions of the remaining layers all adopt a linear L initial function to accelerate the network convergence speed.

5. The method of claim 1, wherein the mean square error function in (3b) is expressed as follows:

wherein W and b represent the weight and bias in the model CE-Net parameter, | |)₂Is the Euclidean norm, T is the number of samples in the training set, H_tAndrespectively the t-th input in the training setThe corresponding desired output and actual output, expressed as:in the formula f_modelRepresenting the channel estimation model CE-Net, W and b represent the weights and offsets in the model parameters, respectively.

6. The method of claim 1, wherein the step (4) of selecting a sample set of the frequency domain signal of the pilot position of the received signal and the corresponding channel vector from the real environment as the target domain data comprises:

according to the similarity measurement method, the similarity between the frequency domain signal of the pilot frequency position of the received signal in the real environment and the frequency domain signal of the pilot frequency position of the received signal in the simulation environment is calculated, and the frequency domain signal of the pilot frequency position of the received signal selected in the real environment with the similarity higher than a set threshold value and the corresponding channel vector are selected to form a sample set as target domain data.

Technical Field

The invention belongs to the technical field of communication, and particularly relates to a channel estimation method in an Orthogonal Frequency Division Multiplexing (OFDM) system, which can be used for the OFDM communication system based on comb-shaped pilot frequency.

Background

OFDM is one of the key technologies widely used in current communication systems, and has low implementation complexity, and can effectively improve the utilization rate of a frequency band. In a broadband mobile communication system, a wireless channel generally has frequency selectivity and time-varying characteristics, and the performance of channel estimation directly affects the quality of a received signal, so that it is necessary to perform dynamic channel estimation and ensure the accuracy of an estimation result.

In the prior research on OFDM channel estimation, the channel estimation principle of the pilot position is mainly focused on the least square L S algorithm and the MMSE algorithm, and the channel interpolation is mainly focused on linear interpolation, Gaussian interpolation and spline interpolation.

When the method is applied to an actual scene, the channel estimation method based on the two principles is difficult to balance between implementation complexity and performance, and the interpolation algorithm cannot track channel changes, so that how to effectively improve the existing channel estimation method and make the method better adapt to engineering application is a problem which needs to be considered emphatically.

In view of these problems, the paper "OFDM channel estimation based on deep learning" proposes a channel estimation method for OFDM system under comb pilot mode, which constructs a deep neural network architecture D L-CE through a fully connected layer neural network module to simulate a channel interpolation process, iteratively trains a channel estimation network through offline channel data, learns the characteristics of the channel, and tracks the channel variation, so that the scheme achieves better performance improvement, but the scheme uses a channel frequency domain response CFR based on L S algorithm to obtain a pilot location as an input of D L-CE, ignores that the CFR obtained based on L S is susceptible to noise, and separates the two processes of channel estimation and channel interpolation of the pilot location, does not consider the linear relationship existing between the received pilot signal and the transmitted pilot signal, increases the algorithm complexity, so that the scheme still has performance deficiency and high algorithm complexity, and thus the scheme has the problem of applying an artificial auxiliary OFDM receiver "with a linear relationship between the received pilot signal and transmitted pilot signal, and the scheme is a solution for fast channel estimation, which is applicable to the conventional adaptive channel estimation, and thus the adaptive to the fast fading of the OFDM channel estimation.

Disclosure of Invention

The invention aims to provide an OFDM channel estimation method based on a full-connection deep neural network FC-DNN (fiber channel-discrete network), aiming at the defects in the prior art, which improves the channel estimation precision on the premise of ensuring lower algorithm complexity and can be applied to a fast fading channel scene.

The technical idea of the invention is as follows: and establishing a channel estimation model of the full-connection deep neural network FC-DNN, namely CE-Net, through the full-connection layer FC neural network. Through data training of the model CE-Net, the mean square error between the output corresponding to the input of the model and the label data corresponding to the input is minimized; the trained channel estimation model CE-Net is deployed at the transmitting end or the receiving end for on-line testing.

According to the technical thought, the OFDM channel estimation method based on deep learning comprises the following steps:

(1) receiving the time domain signal Y by a receiving end and preprocessing the time domain signal Y to obtain a frequency domain signal Y of a pilot frequency position of the received signal_P；

(2) And (3) building a channel estimation model CE-Net by utilizing a full connection layer neural network:

(2a) the channel estimation model CE-Net comprises an input layer, a plurality of hidden layers and an output layer;

(2b) frequency domain signal Y of pilot frequency position of received signal_PAs input to the model, the output of the model is to the true channelEstimated vector of quantity H

(3) Training a channel estimation model CE-Net by using sample data to obtain model parameters:

(3a) generating an input sample set consisting of time domain receiving signals Y and a label sample set consisting of corresponding real channel vectors H by using a Matlab software simulation platform, and preprocessing the input sample set according to the mode described in the step (1) to obtain frequency domain signals Y of pilot frequency positions of the receiving signals_PAnd the sample data composed of the corresponding real channel vector H;

(3b) vector estimation of channelTaking a mean square error function between the real channel vector H and the channel estimation model CE-Net as a cost function of the channel estimation model CE-Net, and performing offline training on the model by using sample data to minimize the cost function to obtain model parameters, namely weight W and bias b; w, b is input into the model to obtain a trained channel estimation model CE-Net;

(4) taking a sample set consisting of a frequency domain signal of a pilot frequency position of a received signal selected from a real environment and a corresponding channel vector as target domain data, carrying out migration training on a trained channel estimation model CE-Net, and adjusting parameters of the channel estimation model CE-Net;

(5) estimating an on-line channel by using the channel estimation model CE-Net obtained in the step (4):

(5a) placing the channel estimation model obtained in the step (4) at a receiving end;

(5b) in the on-line test or use process, the receiving end obtains the frequency domain signal Y of a certain receiving signal pilot frequency position according to the mode described in the step (1)_PInput to the model to obtain a channel estimation vector

Further, the receiving terminal of (1)Receiving and preprocessing the time domain signal Y and obtaining a frequency domain signal Y of the pilot frequency position of the received signal_PIt is implemented as follows:

(1a) the receiving end obtains a time domain receiving signal y, and the receiving end removes a cyclic prefix CP and a discrete Fourier transform DFT from the time domain receiving signal y in sequence to obtain a frequency domain receiving signalWherein C represents a set of complex numbers, N_cThe frequency domain received signal Y contains information Y of pilot frequency position for the number of sub-carriers_PAnd information Y of data position_D；

(1b) Let the number of pilots be N_P＝N_c/D_PIn which N is_cIs the number of subcarriers, D_PIs a pilot interval; because the pilot frequency position and the pilot frequency value are known in advance by the sending end and the receiving end, and the pilot frequency values are all complex v, the frequency domain receiving signal of the pilot frequency position is extracted from the frequency domain receiving signal Y according to the known pilot frequency position

Further, the structure of the channel estimation model CE-Net in (2a) is as follows:

the channel estimation model CE-Net comprises an input layer, a plurality of hidden layers and an output layer;

the input layer and the first hidden layer are both 2N_PA neuron of which N is_PIs the number of pilot frequencies; the output layer is composed of 2N_cA neuron of which N is_cIs the number of subcarriers.

Furthermore, the activation function of the first hidden layer adopts a Sigmoid function, and the activation functions of the other layers all adopt linear L initial functions to accelerate the network convergence speed.

Further, the mean square error function in (3b) is expressed as follows:

wherein W and bRepresenting the weights and biases in the model CE-Net parameters, | |)₂Is the Euclidean norm, T is the number of samples in the training set, H_tAndrespectively the t-th input in the training setThe corresponding desired output and actual output, expressed as:in the formula f_modelRepresenting the channel estimation model CE-Net, W and b represent the weights and offsets in the model parameters, respectively.

Further, the (4) selecting a sample set composed of the frequency domain signal of the pilot position of the received signal and the corresponding channel vector from the real environment as the target domain data includes:

according to the similarity measurement method, the similarity between the frequency domain signal of the pilot frequency position of the received signal in the real environment and the frequency domain signal of the pilot frequency position of the received signal in the simulation environment is calculated, and the frequency domain signal of the pilot frequency position of the received signal selected in the real environment with the similarity higher than a set threshold value and the corresponding channel vector are selected to form a sample set as target domain data.

Compared with the prior art, the invention has the following beneficial effects:

1. compared with a channel estimation technology based on D L-CE, the CE-Net has lower calculation complexity and can obtain higher channel estimation quality under the condition of the same signal to noise ratio, compared with a channel estimation technology based on a man-made fixed interpolation algorithm based on L S, the CE-Net can obtain smaller estimation error under the condition of the same signal to noise ratio, thereby obviously improving the quality of channel estimation.

2. The invention migrates the knowledge obtained by the simulation environment data training to the real environment through the migration learning, and finely adjusts the network model through the data in the real environment, thereby further improving the channel estimation quality of the channel estimation model.

Drawings

FIG. 1 is a block diagram of an implementation flow of the present invention;

fig. 2 is a diagram of a comb pilot structure used in the present invention;

FIG. 3 is an exemplary diagram of a CE-Net network architecture;

fig. 4 is a graph of CE-Net versus quality for several prior art channel estimates.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The invention provides an OFDM channel estimation method based on deep learning. FIG. 1 is a block diagram of an implementation process of the present invention. The following description will be made by way of specific examples.