阅读说明：本技术 深度学习的压缩感知大规模mimo信道估计方法及系统 (Deep learning compressed sensing large-scale MIMO channel estimation method and system ) 是由 何怡刚 黄源 何鎏璐 王枭 程彤彤 隋永波 于 2021-07-08 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压缩感知FDD大规模MIMO系统稀疏信道估计方法及系统,属于无线通信系统导频辅助的信道估计技术领域,其中,方法的实现包括：基于深度学习,即卷积重构网络ConCSNet,在不需要稀疏度的情况下,本发明通过数据驱动的方式,利用ConCSNet求解测量向量y到信号h的逆变换过程,从而解决压缩感知框架下的欠定最优化问题,并实现对原始稀疏信道的重构。仿真结果表明,通过本发明能更快速、准确地恢复稀疏度未知的大规模MIMO系统的信道状态信息。(The invention discloses a compressed sensing FDD large-scale MIMO system sparse channel estimation method and system based on deep learning, belonging to the technical field of pilot frequency assisted channel estimation of a wireless communication system, wherein the method comprises the following steps: based on deep learning, namely a convolution reconstruction network ConCSNet, under the condition that sparsity is not needed, the method solves the inverse transformation process from a measurement vector y to a signal h by using the ConCSNet in a data driving mode, thereby solving the underdetermined optimization problem under a compressed sensing framework and realizing the reconstruction of an original sparse channel. Simulation results show that the method can more quickly and accurately recover the channel state information of the large-scale MIMO system with unknown sparsity.)

1. A compressed sensing FDD large-scale MIMO system sparse channel estimation method based on deep learning is characterized by comprising the following steps:

aiming at an FDD large-scale MIMO downlink system, establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix;

and solving the channel model by adopting a compressive sensing sparse channel estimation ConCSNet network based on deep learning according to the channel model to obtain a sparse signal estimation value.

2. The method of claim 1, wherein the channel model is: y is Th + z, wherein,in order to receive the matrix of signals,in order to be a matrix of channels,in order to be a pilot matrix, the pilot matrix,is a noise matrix, N_RFor the number of receive antennas, W is the number of pilot symbols, N_TFor the number of transmit antennas, L is the channel length.

3. The method of claim 2, wherein the ConCSNet network comprises: the convolutional neural network comprises a linear mapping network and a convolutional neural network behind the linear mapping network, wherein the linear mapping network is a layer of fully-connected layer, and the convolutional neural network is a plurality of convolutional layers which are connected in sequence.

4. The method of claim 3, wherein solving the channel model to obtain sparse signal estimation values according to the channel model by using a deep learning based compressed sensing sparse channel estimation ConCSNet comprises:

the sparse channel matrix is compressed and sampled by the perception pilot frequency matrix to obtain a receiving signal matrix, and the complex number is converted into a complex numberInput matrix converted into ConCSNet networkWherein, y_RIs the real part of y, y_IIs the imaginary part of the y, and,is the value of the energy of y,are respectively a vector y_EElement of (1), and a tagInto a three-dimensional N_RN_TX L x 2 matrixThe 3 rd dimension of the matrix is used for describing a complex part and an imaginary part of data, then a linear mapping network of a full connection layer is used for obtaining a first reconstruction channel of the channel, and a convolutional neural network of a multi-layer convolutional layer is used for obtaining a second reconstruction channel.

5. The method of claim 4, wherein the linear mapping network has a loss function of:wherein, F^f(. represents a linear mapping, y_iFor the measured value of the received signal matrix corresponding to the i-th training sample, h_iIs the time domain channel information of the channel matrix corresponding to the ith training sample, i is 1,2^fIs a linear mapping matrix.

6. The method of claim 5, wherein the method is performed byA first reconstructed channel of the channel is obtained using a linear mapping network of the full link layer, wherein,channel information of a first reconstructed channel obtained for a linear mapping process.

7. The method of claim 6, wherein the loss function of the convolutional neural network is:wherein omega^cAs a parameter of a convolutional neural network, F^c(. cndot.) represents a convolutional mapping process.

8. The method of claim 7, wherein the method is performed byObtaining a second reconstructed channel from the convolutional neural network of the multi-layered convolutional layer, wherein,channel information of the second reconstructed channel obtained for the convolutional neural network,the real and imaginary parts of the data are then combined and further converted to a complex channel matrix

9. A compressed sensing FDD massive MIMO system sparse channel estimation system based on deep learning is characterized by comprising the following steps:

the channel model establishing module is used for establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix aiming at an FDD large-scale MIMO downlink system;

and the channel estimation module is used for solving the channel model by adopting a compressive sensing sparse channel estimation ConCSNet network based on deep learning according to the channel model to obtain a sparse signal estimation value.

10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 8.

Technical Field

The invention belongs to the technical field of pilot frequency assisted channel estimation of a wireless communication system, and particularly relates to a compressed sensing sparse channel estimation method and system based on deep learning, namely compressed sensing sparse channel estimation based on a convolution reconstruction network ConCSNet.

Background

Massive MIMO technology is considered as one of promising technologies in future wireless communication due to the advantages of high spectrum utilization and large beamforming gain. In the downlink of a Frequency Division Duplex (FDD) massive MIMO system, accurate channel state information is crucial to beamforming. With the rapid increase of the number of antennas at the base station end, accurate estimation of FDD massive MIMO sparse channels is a challenging problem with low pilot overhead.

In a large-scale MIMO system, channel measurement finds that the time domain, the frequency domain and the space domain of a wireless multipath channel have sparse characteristics. By exploiting the sparseness of the channel, implementing an effective channel estimation method by applying a Compressed Sensing (CS) correlation theory has become a hot spot of current research. For Downlink Channel Estimation of a large-scale MIMO system in FDD mode, reference is made to "Low-Complexity Downlink Channel Estimation for Millimeter-Wave FDD Massive MIMO Systems" (Wu X, Yang G, Hou F, et al. ieee Wireless Communication receivers, 2019,8(4):1103 and 1107.), based on the inherent time domain sparsity of a user Channel matrix, an improved Orthogonal Matching Pursuit (OMP) algorithm is proposed to reduce pilot overhead and improve Channel Estimation accuracy. In order to get rid of the dependence of the OMP algorithm on the prior condition of the channel Sparsity, reference "Estimation of Block Sparse Channels with joint Gradient SAMP in Massive MIMO Systems" (Wang P, Zhang H, Yang l.20194th International Conference on electrochemical Control Technology and transport (icect). 2019.) further proposes a Block Sparse adaptive matching pursuit (BCG-SAMP) algorithm, which takes into account the spatio-temporal Sparsity of the delay domain Massive MIMO channel and is able to adaptively obtain the Sparsity level of the channel structured Sparse matrix. Although the accuracy of the BCG-SAMP algorithm is high, the iteration process of the algorithm is limited by the estimation quantity parameters of channel noise, and the estimation performance of the algorithm is rapidly reduced in the presence of noise. However, most of these methods adopt structured sparse assumptions to estimate the wireless channel and use an iterative optimization strategy to solve the underdetermined optimization problem. The intensive calculation of the iterative optimization becomes a bottleneck of the CS in the wireless channel estimation application, so how to further increase the operation speed of the compressed sensing reconstruction algorithm becomes a difficult problem.

Disclosure of Invention

Aiming at the defects or improvement requirements in the prior art, the invention provides a compressive sensing sparse channel estimation method and system based on a deep learning convolution reconstruction network ConCSNet, aiming at solving the problem of sparse channel estimation of the existing FDD downlink massive MIMO system and quickly and accurately recovering massive MIMO channel information with unknown sparsity.

In order to achieve the above object, according to an aspect of the present invention, there is provided a compressed sensing FDD large-scale MIMO system sparse channel estimation method based on deep learning, including:

aiming at an FDD large-scale MIMO downlink system, establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix;

and solving the channel model by adopting a compressive sensing sparse channel estimation ConCSNet network based on deep learning according to the channel model to obtain a sparse signal estimation value.

In some optional embodiments, the channel model is: y is Th + z, wherein,in order to receive the matrix of signals,in order to be a matrix of channels,in order to be a pilot matrix, the pilot matrix,is a noise matrix, N_RFor the number of receive antennas, W is the number of pilot symbols, N_TFor the number of transmit antennas, L is the channel length.

In some optional embodiments, the ConCSNet network comprises: the convolutional neural network comprises a linear mapping network and a convolutional neural network behind the linear mapping network, wherein the linear mapping network is a layer of fully-connected layer, and the convolutional neural network is a plurality of convolutional layers which are connected in sequence.

In some optional embodiments, solving the channel model to obtain a sparse signal estimation value by using a deep learning-based compressed sensing sparse channel estimation concsn according to the channel model includes:

the sparse channel matrix is compressed and sampled by the perception pilot frequency matrix to obtain a receiving signal matrix, and the complex number is converted into a complex numberInput matrix converted into ConCSNet networkWherein, y_RIs the real part of y, y_IIs the imaginary part of the y, and,is the energy value of y, y₁，…，Are respectively a vector y_EElement of (1), and a tagInto a three-dimensional N_RN_TX L x 2 matrixWherein the 3 rd dimension of the matrix is used to characterize the complex and imaginary parts of the data, and then a first reconstructed channel of the channel is obtained using a fully-connected layer linear mapping network, and the first reconstructed channel is derived from the first reconstructed channelThe convolutional neural network of the multilayer convolutional layer obtains a second reconstructed channel.

In some alternative embodiments, the linear mapping network has a loss function of:wherein, F^f(. represents a linear mapping, y_iFor the measured value of the received signal matrix corresponding to the i-th training sample, h_iIs the time domain channel information of the channel matrix corresponding to the ith training sample, i is 1,2^fIs a linear mapping matrix.

In some alternative embodiments, the composition is prepared byA first reconstructed channel of the channel is obtained using a linear mapping network of the full link layer, wherein,channel information of a first reconstructed channel obtained for a linear mapping process.

In some alternative embodiments, the loss function of the convolutional neural network is:wherein omega^cAs a parameter of a convolutional neural network, F^c(. cndot.) represents a convolutional mapping process.

In some alternative embodiments, the composition is prepared byObtaining a second reconstructed channel from the convolutional neural network of the multi-layered convolutional layer, wherein,channel information of the second reconstructed channel obtained for the convolutional neural network,the real and imaginary parts of the data are then combined and further converted to a complex channel matrix

According to another aspect of the present invention, there is provided a sparse channel estimation system for a deep learning-based compressed sensing FDD massive MIMO system, including:

the channel model establishing module is used for establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix aiming at an FDD large-scale MIMO downlink system;

and the channel estimation module is used for solving the channel model by adopting a compressive sensing sparse channel estimation ConCSNet network based on deep learning according to the channel model to obtain a sparse signal estimation value.

According to another aspect of the invention, a computer-readable storage medium is provided, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method of any of the above.

In general, compared with the prior art, the above technical solution contemplated by the present invention can achieve the following beneficial effects:

the invention provides a novel ConCSNet method for estimating a compressed sensing sparse channel based on deep learning, which is oriented to an FDD large-scale MIMO downlink system model. Under the condition that sparsity is not needed, the inverse transformation process from the vector y to the signal h is solved by a data driving mode and ConCSNet, so that the underdetermined optimization problem under a CS framework is solved, and the reconstruction of an original channel is realized. Experimental results show that compared with the traditional CS-based channel estimation algorithm, the ConCSNet method provided by the invention has the advantages that the performance is obviously improved, and the reconstruction speed can be improved by 2-3 times. The problems of long calculation time and poor reconstruction effect of the traditional CS channel estimation algorithm based on iteration are solved.

Drawings

Fig. 1 is a schematic flowchart of a compressed sensing channel estimation method based on deep learning according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a compressed sensing channel estimation method based on deep learning according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of a concsunt framework according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating a comparison of the MSE performance of the channel estimation in each method according to the embodiment of the present invention and the comparative embodiment at different SNR;

fig. 5 is a schematic diagram illustrating comparison of channel estimation MSE performance of different methods according to different sampling rates in the embodiment and the comparative embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is 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. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

In the present examples, "first", "second", etc. are used for distinguishing different objects, and are not used for describing a specific order or sequence.

The FDD massive MIMO system sparse channel estimation technology based on compressed sensing has been widely researched and applied in recent years, but two weaknesses of the technology are gradually highlighted: first, it is difficult to accurately obtain a priori knowledge of channel sparsity. Second, the existing reconstruction algorithms have a slow convergence rate, thus limiting the CS technique to applications in non-real-time scenarios. For FDD massive MIMO downlink system, in this embodiment, the performance of the proposed compressed Sensing sparse channel estimation Convolutional reconstruction Network (con cenet) method based on deep learning will be verified. The number of the antennas at the transmitting end and the receiving end of the simulation system is 16 and 4 respectively; the total number of subcarriers of an Orthogonal Frequency Division Multiplexing (OFDM) technology is 1024; the length and sparsity of each path channel are 128 and 9 respectively; the number of pilots is 64 and all the pilotsThe frequencies are randomly placed in a block-like manner. In this embodiment, an LTE-Advanced channel model is adopted, which mainly includes three application scenarios, including Extended Pedestrian A (EPA), Extended Vehicular A (EVA), and Extended Typical Urban model (ETU). In this embodiment, an EVA wireless communication environment is mainly considered. The number of paths S is 6, the path delays are 0,30,150,310,370,710,1090,1730,2510(ns), and the path gains are-3, 0, -2, -6, -8, -10(dB), respectively. Based on the sparsity and compressive sensing framework of massive MIMO channel, consider a single-cell multi-user FDD massive MIMO system downlink system with N as shown in FIG. 1_TRoot transmitting antenna, N_RThe root receives the antenna and the N subcarriers. Assuming that the channel parameters are constant in one OFDM symbol and the channel length is L, the system has W pilot symbols and is respectively located at the sub-carriers k₁,k₂,…,k_WThe above. Based on the sparsity and compressive sensing framework of a large-scale MIMO channel, the compressive sensing sparse channel estimation ConCSNet method based on deep learning provided by the invention comprises the following steps:

s1: aiming at an FDD large-scale MIMO downlink system, establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix;

in this embodiment, the established channel model is:

y＝Th+z (1)

wherein the content of the first and second substances,in order to receive the matrix of signals,in order to be a matrix of channels,in order to be a pilot matrix, the pilot matrix,is a noise matrix. Due to the fact thatN_RW＜＜N_RN_TL, the channel model is an underdetermined equation, but a joint sparsity structure exists in a large-scale MIMO channel, and a high-dimensional channel h can be reconstructed from a low-dimensional vector y by the compressed sensing sparse channel estimation method based on deep learning.

In the present embodiment, MIMO refers to multiple input multiple output, and standard MIMO networks will typically use 2 or 4 antenna configurations. On the other hand, massive MIMO is a MIMO system that uses a large number of antennas. As for the construction setup of massive MIMO, there is no current debate, but a system with tens or even hundreds of antennas can be called massive MIMO in general description.

S2: solving sparse signal estimation values by using deep learning-based compressed sensing sparse channel estimation ConCSNet according to a channel model

In the present embodiment, as shown in fig. 2, step S2 can be implemented by:

the sparse channel h is compressed and sampled through the sensing matrix T to obtain a received signal y, then an approximate solution of the channel is obtained by using a linear mapping network of a full connection layer, and a reconstructed channel with higher quality is obtained by using a multilayer convolutional neural network.

The approximate solution in this embodiment is to obtain a first reconstructed channel of a channel using a linear mapping network of a fully-connected layer, and the reconstructed channel with higher quality is to obtain a second reconstructed channel using a convolutional neural network of a multi-layer convolutional layer.

Further, it is noted that the complexity of reconstruction increases due to the noise of the wireless communication system and the fact that the received signal y and the channel information h are both complex. At the input of ConCSNet, the complex numberConversion into an input matrixWherein y is_RIs the real part of y, y_IIs the imaginary part of the y, and,is the energy value of y, y₁，…，Are respectively a vector y_EOf (1). While the labelInto a three-dimensional N_RN_TX L x 2 matrixWhere the 3 rd dimension of the matrix is used to delineate the complex and imaginary parts of the data. Secondly, the ConCSNet algorithm increases the number of convolutional layers on one hand, expands the scale of the convolutional neural network, and reduces part of larger convolutional kernels on the other hand, so that the network focuses more on reconstruction details. As shown in fig. 3, the linear mapping network and the convolutional neural network included in the ConCSNet network are as follows:

(1) linear mapping network

For equation (1), the original channel information h is reconstructed from the measurement signal y, and this process can be approximately regarded as a linear mapping, i.e., h ═ Ω y. Wherein the content of the first and second substances,is a linear mapping matrix. Due to N_RW＜＜N_RN_TL, the mapping process solves an underdetermined optimization problem, and an accurate solution is difficult to solve. The embodiment adopts a linear mapping network F^fObtaining initial reconstructed channel, channel information obtained in linear mapping processTo approximate the solution, the corresponding linear mapping matrix is Ω^fAnd makeThe error of (2) is minimal. Assume that the training set contains M training samples, i.e., { (y)₁,h₁),(y₂,h₂),…,(y_M,h_M)}. Wherein, y_iFor measuring received signals, h_iFor time domain channel information, i is 1, 2. For the one-layer full-connection layer linear mapping network, the loss function is as follows:

wherein, F^f(. -) represents a linear mapping, and the training process is trained using the Adam (Adaptive motion Estimation, Adam) method. The network gets an approximate solution for channel h as:

(2) convolutional neural network

For the linear mapping network in (1), an approximate solution of the estimated channel can be obtainedIn order to obtain a high-precision channel estimation value, further processing is required, and in this embodiment, a full convolutional network consisting of 9 convolutional layers is added after the linear mapping network. The relevant description is as follows: the first layer and the fifth layer both use 7 × 7 convolution kernels and generate 64 feature maps; the second layer and the sixth layer use 5 × 5 convolution kernels, and 32 feature maps are generated; the third layer and the seventh layer both use a convolution kernel of 3 multiplied by 3 and generate 16 feature maps; the fourth layer and the eighth layer both use 1 × 1 convolution kernels and generate 2 feature maps; the ninth layer generates 2 feature maps using a 1 × 1 convolution kernel and outputs a high-precision channel reconstruction result. The activation function is used by all eight layers except the ninth layer, which does not need the ReLU activation function. The entire network uses all-1 padding to ensure that the size of the feature map in all layers remains unchanged. The parameter omega of the full connection layer obtained by the formula (2)^fAs an initial value of convolutional neural network training, and using Adam optimization algorithm to update parameter omega in linear mapping matrix^fAnd parameter omega of convolutional neural network^c. During training, equation (4) is used as a loss function, namely:

the exact solution of channel h obtained by the convolutional learning network is:

wherein, F^c(·) denotes a convolutional mapping process in which,the real and imaginary parts of the data are then combined and further converted to a complex channel matrix

(3) ConCSNet network training parameter configuration

In order to train the ConCSNet network proposed by the present invention, the training process is divided into two steps. First, the linear mapping network F is trained^f(. cndot.), learning rate 0.001, momentum factor 0.95. F^fAfter the network training is finished, the whole ConCSNet network is trained, and the small learning rate is 0.0001, and the Adam method is also used for training, wherein the momentum factor is 0.99. Second, the training, validation and test data sets contained 128000, 38400 and 38400 samples, respectively. The batch size is 128, the learning rate decay factor is 0.96, and the decay period of the learning rate is 100. The present invention implements the network using a tensoflow framework and trains it at 2.8GHz using an Intel Core i5-4200H CPU.

In order to further objectively evaluate the performance of the proposed ConCSNet deep learning channel estimation method, a Mean Square Error (MSE) is used as an evaluation index for analysis. Wherein, MSE is used for measuring the difference between the reconstruction value and the real value, and the smaller the MSE is, the better the reconstruction performance is. The expression for MSE is:

wherein N is_EFor the simulation times, the value is 20.

Figure 4 shows the MSE performance of several methods at different signal-to-noise ratios. Experimental results show that SAMP has gradually poor reconstruction performance under low signal-to-noise ratio, and the reconstruction performance is superior to that of the gOMP method under higher signal-to-noise ratio. The ConCSNet and Reconnet reconstruction methods represented by deep learning have better reconstruction performance than traditional iterative compressed sensing reconstruction algorithms (OMP, SAMP, CoSaMP, gOMP) under each signal-to-noise ratio. The ConCSNet method provided by the invention optimizes the input information, the structure of the convolutional layer and the number of layers, so that the reconstruction performance is further improved compared with Reconnet.

Fig. 5 shows the MSE and PSNR performance of several methods at different sampling rates, respectively. Experimental results show that except LS methods, the reconstruction performance of each method can be improved by increasing the sampling rate. At lower sampling rates, the performance of the OMP method is progressively worse than the LS, while several other reconstruction methods (SAMP, CoSaMP, giomp) also approach the LS method. This indicates that the conventional iterative compressed sensing reconstruction method cannot effectively reconstruct the channel at low signal-to-noise ratio. And ConCSNet and Reconnet reconstruction methods based on deep learning can keep better performance, and ConCSNet is also better than Reconnet.

The ConCSNet method for sparse channel estimation of the FDD large-scale MIMO system based on the deep learning compressed sensing has the characteristic of good reconstruction performance, and is suitable for occasions needing pilot frequency auxiliary channel estimation of a wireless communication system.

In another embodiment of the present invention, a compressed sensing FDD massive MIMO system sparse channel estimation system based on deep learning is further provided, including:

the channel model establishing module is used for establishing a channel model based on a channel matrix, a pilot matrix, a received signal matrix and a noise matrix aiming at an FDD large-scale MIMO downlink system;

and the channel estimation module is used for solving the channel model by adopting a compressive sensing sparse channel estimation ConCSNet network based on deep learning according to the channel model to obtain a sparse signal estimation value.

The specific implementation of each module may refer to the description of the above method embodiment, and this embodiment will not be repeated.

It should be noted that, according to the implementation requirement, each step/component described in the present application can be divided into more steps/components, and two or more steps/components or partial operations of the steps/components can be combined into new steps/components to achieve the purpose of the present invention.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.