WO2020253690A1

WO2020253690A1 - Deep learning beam domain channel estimation method based on approximate message passing algorithm

Info

Publication number: WO2020253690A1
Application number: PCT/CN2020/096428
Authority: WO
Inventors: 韦逸; 赵明敏; 赵民建; 雷鸣
Original assignee: 浙江大学
Priority date: 2019-06-17
Filing date: 2020-06-16
Publication date: 2020-12-24
Also published as: CN110365612B; CN110365612A

Abstract

Provided are a deep learning beam domain channel estimation method based on an approximate message passing algorithm, wherein same is mainly used in a millimeter wave massive MIMO system based on a lens antenna. The method comprises the following steps: (1) constructing a deep network structure, wherein the deep network is mainly composed of two parts: one is a model-driven deep network LAMP based on an approximate message passing algorithm, and the other is a data-driven deep network ResNet based on residual learning; (2) modeling a beam domain channel according to the geometric structure of a lens antenna, and generating training data according to a system model; (3) using training data with different signal-to-noise ratios to train a network offline; and (4) fixing the optimized network parameters, and using the trained network to perform real-time beam domain channel estimation according to a received signal at a radio frequency chain end. The present invention can effectively improve the accuracy of beam domain channel estimation, and possesses a similar computational complexity to that of a traditional channel estimation algorithm.

Description

A Channel Estimation Method Based on Approximate Message Passing Algorithm in Deep Learning Beam Domain

Technical field

The invention belongs to the field of wireless communication, and is a deep learning beam domain channel estimation method based on an approximate message transfer algorithm.

Background technique

With the explosive growth of mobile data demand, the fifth-generation (5G) communication network uses the rich spectrum resources of the millimeter wave band to increase communication capacity. However, millimeter wave communication has the defect of large in-band penetration loss, which can cause severe channel fading. Millimeter-wave massive MIMO systems can use large antenna arrays to provide high data rates to compensate for in-band penetration loss.

However, when each antenna is equipped with a radio frequency chain, the realization of the millimeter-wave massive MIMO antenna system will be accompanied by unaffordable hardware complexity and power consumption. An effective way to reduce the complexity of implementation is to use advanced lens antenna arrays. The lens antenna can convert the traditional space-domain channel into a beam-domain channel, playing the role of a beam-domain separation Fourier transform (DFT) matrix. Due to insufficient scattering effects in millimeter wave frequencies, the number of effective propagation paths is limited. Therefore, the beam domain channel has the characteristic of sparseness, and we can reduce the number of RF chains by selecting the main beam.

When the dimension of the channel matrix is greater than the number of RF chains, the beam-domain channel estimation can be regarded as a sparse signal recovery problem. Many channel estimation algorithms based on compressed sensing have been widely used, such as channel estimation algorithm AMP based on approximate message passing algorithm, channel estimation algorithm CoSaMP based on compressed sampling matching tracking, channel estimation algorithm StOMP based on segmented orthogonal matching tracking, etc. .

In addition, since the deep learning method has been successfully applied in many other fields, such as image processing, natural language processing, etc., it has also begun to be used in the field of wireless communication as a potential technology, such as signal detection and channel estimation. There are two mainstream deep learning methods. One is the model-driven deep learning method, which builds networks based on known knowledge and mechanisms; the other is the data-driven deep learning method, which regards the network as a black box and relies on a large number of Data training this network, common fully connected networks and deep convolutional networks belong to this method. The present invention combines the advantages of the above two methods, and the adopted deep network is mainly composed of two parts: One is the model-driven deep network LAMP based on the approximate message passing algorithm, and this sub-network inherits the sparse recovery capability of the approximate message passing algorithm, A preliminary estimation result can be obtained; the second is the data-driven deep network ResNet based on residual learning, which can further eliminate the residual error between the beam domain channel matrix and its estimated value based on the preliminary estimation result, and reduce noise Influence, obtain more accurate channel estimation results. The network is trained offline from simulation or measured data to obtain network parameters. After the training is completed, online real-time channel estimation is completed through fixed parameters. This method effectively improves the accuracy of channel estimation while maintaining the computational complexity of traditional channel estimation methods.

Summary of the invention

The purpose of the present invention is that in a millimeter wave massive MIMO system based on a lens antenna, it is difficult for traditional algorithms to estimate a high-dimensional beam domain channel matrix from a low-dimensional received signal obtained from a limited radio frequency chain end, and a method based on approximate message transmission is proposed. Algorithmic deep learning beam domain channel estimation method, the present invention adopts the following technical solutions:

The application scenario of the present invention is a millimeter wave massive MIMO system based on a lens antenna, and the system model is shown in FIG. 1. As the receiving end, the base station is equipped with a three-dimensional electromagnetic lens, and an antenna array with a scale of N _rz × N _ry is placed on its focal plane. The N _r =N _rz ×N _ry antennas are connected to M radio frequency chains through a selection network W with a scale of M×N _r , and the number of radio frequency chains is smaller than the number of antennas. The selection network W is composed of M columns extracted from a randomly generated N _r ×N _r Bernoulli matrix.

The specific steps of the present invention are as follows:

1. Construct a deep learning network structure. The deep learning network structure consists of two parts: model-driven deep sub-network LAMP based on approximate message passing algorithm and data-driven deep sub-network ResNet based on residual learning; model-driven deep sub-network LAMP The output of the previous stage is used as the input of the latter stage of the data-driven deep sub-network ResNet;

2. Model the beam domain channel according to the geometric parameters of the lens antenna, obtain the beam domain channel matrix, and obtain training data with different signal-to-noise ratios according to the MIMO system model;

3. Use the training data with different signal-to-noise ratios described in step 2 to perform offline training on the model-driven deep sub-network LAMP to obtain the network parameters of the model-driven deep sub-network LAMP; fix the network parameters and compare the parameters described in step 1. Perform end-to-end training on the deep learning network structure of, and get the deep learning network model after training;

4. According to the received signal at the RF chain end, use the trained deep learning network model to perform real-time beam domain channel estimation.

Further, the step 1 is specifically:

1.1. Construct a model-driven deep sub-network LAMP based on an approximate message passing algorithm. The model-driven deep sub-network LAMP is developed by an approximate message passing algorithm. It is composed of T layers and each layer has the same structure. The i iteration is represented as the i-th layer of the model-driven deep sub-network LAMP, the parameters (W, W ^H ) contained in the t-th iteration are replaced by the learnable parameters (β _t W, Z _t ) of the t-th layer, and W is The radio frequency chain selection network of the base station; the beam domain channel estimation process of the model-driven deep sub-network LAMP layer t is expressed as:

Where r is the received signal,

Is the beam domain estimation channel of the t-th layer, v _t is an intermediate variable, M is the number of receiving antennas, α _t is a network parameter, and η(·; λ) is a soft threshold function, which is defined as follows:

[η(u;λ)] _j =sgn(u _j )max{|u _j |-λ,0}

Where u _j is the j-th element of an independent variable vector of the function, λ is another independent variable of the function, which represents the threshold, and sgn(·) is a sign function;

The first-stage beam domain estimation channel output by the model-driven deep sub-network LAMP is expressed as:

among them

Represents the mapping of the model-driven deep sub-network LAMP, Θ={β _t ,Z _t ,α _t } represents the learning parameters of the model-driven deep sub-network LAMP, and r ⁱ is the received signal;

The L2 norm is selected as the cost function, and the specific loss function is expressed as:

Where N is the number of training data, h ⁱ is a channel matrix;

1.2. Building a data-driven deep sub-network ResNet based on the residual learning algorithm. The data-driven deep sub-network ResNet is composed of a number of residual blocks with the same structure. Each residual block has multiple convolutional layers, and each layer of volume An activation layer is connected to the back of the accumulation layer. The activation function adopts tanh(·). The activation function can be guided everywhere, and the input variable is mapped to (-1,1);

The output of the data-driven deep sub-network ResNet is expressed as:

among them

In order to obtain the residual error between the estimated channel and its estimated value through ResNet,

Represents the mapping of the data-driven deep sub-network ResNet, Σ is the learning parameter of the data-driven deep sub-network ResNet, that is, the weight and bias in the convolutional layer;

1.3. The received signal at the RF chain end passes through

with

Get the second-level beam domain estimation channel

Further, the step 2 is specifically:

2.1. Model the beam domain channel according to the geometric parameters of the lens antenna, and the beam domain channel model is expressed as:

Where L represents the number of diameters, α ^(l) is the amplitude of the diameter l, φ ^(l) and θ ^(l) represent the incident azimuth and altitude angles of the diameter l respectively, A _r (φ ^(l) ,θ ^(l) ) Is the response matrix of the lens antenna array, which is determined by the geometric parameters of the lens antenna. The (y, z)th element of the response matrix is expressed in the form of the product of two sinc(·) functions:

Where D _Y and D _Z represent the length and height of the lens antenna, and λ is the wavelength of the incident wave;

Obtain the beam domain channel matrix according to the beam domain channel model

2.2. Obtain the received signal r at the RF chain end according to the MIMO system model,

Where s is a known pilot signal,

Is equivalent noise, n is Gaussian white noise, and h is the vectorized beam domain channel matrix;

Change the signal-to-noise ratio of the training data by adjusting the variance of the Gaussian white noise n; the vectorized beam-domain channel matrix

And the corresponding received signal vector at the RF chain end

Form the required training data

N is the number of training data.

Further, the step 3 is specifically as follows:

3.1. Use training data with a high signal-to-noise ratio above 10dB to train the model-driven deep sub-network LAMP layer by layer. When training the t-th layer, the parameters of the first t-1 layer

Keep it unchanged, using a step-down learning rate, the initial value is a, every K times of training, the learning rate is reduced to the original τ, and the training is terminated when the normalized mean square error no longer decreases;

3.2. Use the training data with high signal-to-noise ratio described in step 3.1 for end-to-end training of the model-driven deep sub-network LAMP, with a learning rate of 0.1×a, and the training will be terminated when the normalized mean square error no longer decreases;

3.3. Use the training data with different signal-to-noise ratios described in step 3 to optimize the deep learning network structure end-to-end. The learning rate is 0.01×a. When the normalized mean square error no longer decreases, the training is terminated, and finally the training is obtained. Deep learning network model.

The beneficial effects of the present invention: the present invention fully combines the model-driven deep learning method and the data-driven deep learning method, and utilizes the sparsity characteristics of the beam domain channel matrix, and uses the approximate message propagation algorithm commonly used in the field of sparse signal recovery. The iterative process is expanded into a deep network, and the fixed parameters in the algorithm are transformed into learnable parameters. With the help of deep learning, performance is improved. At the same time, the idea of residual learning is introduced to further reduce the difference between the channel and its estimated value, so as to further improve the accuracy of the estimation result and reduce the ability of the network to resist noise. All training is done offline. Once the training is completed, only one forward calculation is needed according to the received signal to get the estimation result. In the online real-time estimation process, the present invention can obtain a computational complexity similar to that of traditional algorithms such as approximate message passing algorithms, and improves the accuracy of channel estimation without increasing computational complexity.

Description of the drawings

The above and/or additional aspects and advantages of the present application will become obvious and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, in which:

Figure 1 is a millimeter-wave massive MIMO system model based on a lens antenna;

Figure 2 is a block diagram of a single-layer structure of a model-driven deep network based on an approximate message passing algorithm;

Figure 3 is a structure diagram of a deep learning network based on an approximate message passing algorithm;

Figure 4 is the normalized mean square error performance curve of the deep learning channel estimation method based on the approximate message passing algorithm;

Figure 5 is the arrival rate performance curve of the deep learning channel estimation method based on the approximate message passing algorithm.

Detailed ways

In order to make the technical solution and advantages of the present invention clearer, the specific implementation of the technical solution will be described in more detail in conjunction with the accompanying drawings:

The application scenario of this embodiment is a millimeter wave massive MIMO system based on a lens antenna, and the system model is shown in FIG. 1. In the millimeter-wave massive MIMO system based on lens antennas, the base station is used as the receiving end, and a three-dimensional electromagnetic lens is installed, and a 32×32 antenna array is placed on its focal plane. The 1024 antennas are connected to 819 radio frequency chains through a selection network W with a scale of 819×1024, and the number of radio frequency chains is smaller than the number of antennas. The selection network W is composed of M columns extracted from a randomly generated 1024×1024 Bernoulli matrix. The proposed deep learning beam-domain channel estimation method based on the approximate message passing algorithm for this system includes the following steps:

Step 1: Build a deep learning network.

In this embodiment, the deep learning network is introduced into the base station, and the received signal vector at the finite radio frequency link is used as the input signal, and the estimated value of the beam domain channel matrix is output through the forward calculation of the network. The deep learning network used is mainly composed of two parts: one is the model-driven deep network LAMP based on approximate message passing algorithm, and the other is the data-driven deep network ResNet based on residual learning.

The model-driven deep learning network LAMP based on the approximate message passing algorithm is developed by the approximate message passing algorithm. It consists of 5 layers and each layer has the same structure. Each iteration of the approximate message passing algorithm is regarded as a layer of the sub-network, and the (W, W ^H ) contained in the tth iteration are replaced by the learnable parameters (β _t W, Z _t ) based on this layer . As shown in Figure 2, for the t-th layer of the LAMP sub-network, the channel estimation process can be expressed as:

Where r is the received signal,

Is the estimated channel of each layer, and v _t is the intermediate variable. η(·;λ) is a soft threshold function, which is defined as follows:

[η(u;λ)] _j =sgn(u _j )max{|u _j |-λ,0}

The output of the LAMP subnet can be expressed as

Represents the mapping process of the sub-network, Θ={β _t ,Z _t ,α _t } is the parameter to be learned. Use the L2 norm as the loss function, specifically:

The data-driven deep sub-network ResNet based on residual learning protects the integrity of the information by directly detouring the input information to the output. The entire network only needs to learn the difference between input and output

Simplify learning objectives and difficulty. This sub-network can further eliminate the residual error between the beam domain channel matrix and its estimated value based on the preliminary estimation result, reduce the influence of noise, and obtain more accurate channel estimation results.

The ResNet sub-network is composed of multiple Residual Blocks with the same structure, and each Residual Block has three convolutional layers. The first layer uses 7×7 convolution kernels to generate 64 feature mapping layers, the second layer uses 5×5 convolution kernels to generate 32 feature mapping layers, and the third layer uses 3×3 convolution kernels to generate 1 feature mapping layer. . Each layer of convolutional layer is connected with an activation layer, and the activation function uses tanh(·).

The output of this subnet can be expressed as

Is the mapping represented by the residual learning network,

It is the previous output of the model-driven deep sub-network based on the approximate message passing algorithm, as the input of the data-driven deep sub-network based on residual learning, Σ is the learning parameter included in the residual learning network, that is, the weight and sum in the convolutional layer Bias.

As shown in Figure 3, the above two sub-networks constitute the basic structural blocks of the proposed network, and multiple basic structural blocks constitute the entire network. The received signal at the RF chain end passes through

with

, The final estimate of the channel matrix is obtained

It can be expressed as:

Use the L2 norm as the cost function, which is expressed as follows:

The complete network model is shown in Figure 3. In this example, the network is composed of two basic network functional blocks. Each basic functional block contains five layers of LAMP sub-networks and one residual block.

Step 2: Collect the training data set.

In this method, supervised learning is used to optimize the unknown parameters in the network, so a large amount of labeled training data needs to be collected. Model the beam domain channel transformed from the lens antenna according to the geometric parameters of the lens antenna, and obtain a series of beam domain channel matrices

The beam domain channel of the system can be modeled as:

Where L represents the number of diameters, α ^(l) is the amplitude of the diameter l, and φ ^(l) and θ ^(l) represent the incident azimuth and height angle of the diameter l, respectively.

A _r (φ ^(l) ,θ ^(l) ) is the response matrix of the antenna array, which is determined by the geometric characteristics of the lens antenna. The (y,z)th element of the response matrix can be expressed in the form of the product of two sinc(·) functions:

Among them, D _Y and D _Z represent the length and height of the lens antenna, and λ is the wavelength of the incident wave.

A single-antenna user sends a known pilot signal s to the base station with different signal-to-noise ratios, and obtains the received signal r at the RF chain end according to the system model. R can be expressed as:

among them

Is equivalent noise, n is Gaussian white noise.

Vectorized channel matrix

And the corresponding RF chain end received signal vector

Constitute the required data label group

N is the number of training data.

Step three, offline training.

All training processes are completed offline, the training of the network is implemented on the tensorflow platform, and the ADAM trainer is used for training. The whole training process is divided into three steps, and the training is terminated when the normalized mean square error no longer decreases. The first step uses high signal-to-noise ratio training data to train the model-driven deep network LAMP based on the approximate message passing algorithm layer by layer. The specific method is: when training the t-th layer, the parameters of the previous t-1 remain unchanged. And a step-down learning rate is used, the initial value is 0.001, and the learning rate is reduced to the original 0.5 for every 10,000 training sessions. The second step adopts an end-to-end training method, using the high signal-to-noise ratio training data in the first step to perform end-to-end training on the entire network, and the learning rate is 0.0001. The third step uses training data under different signal-to-noise ratios to optimize the end-to-end network as a whole to enhance its anti-interference ability against noise. In this step, the learning rate is 0.00001.

Step 4. Online estimation.

Once the training is over, the parameters in the network are stored for online real-time beam-domain channel estimation. A single antenna user sends the same pilot signal s to the base station, and the received signal vector at the RF link is directly sent to the trained deep network. After a forward operation, the estimated beam domain channel matrix is directly output for subsequent use Signal detection.

Figure 4 shows the estimation accuracy of different channel estimation algorithms under different SNR conditions, measured by the normalized mean square error. LampResNet represents the channel estimation method proposed by the present invention. StOMP, AMP and CoSaMp are three channel estimation algorithms based on compressed sensing, and LAMP, LDAMP and DR2-Net are three contrast algorithms based on deep learning. It can be seen from the figure that the channel estimation method proposed by the present invention obtains the best estimation accuracy under all signal-to-noise ratios.

Figure 5 shows the arrival rate performance under multi-user conditions. The number of users is 2. Each user transmits orthogonal pilot signals to the base station. It is assumed that the channels experienced by each user are different. It can be seen from the figure that the invented method can obtain the maximum arrival rate in comparison with the channel estimation algorithm under multi-user conditions.

The invention is a deep learning beam domain channel estimation method applied to a millimeter wave massive MIMO system based on a lens antenna and based on an approximate message transfer algorithm. Regarding the deep learning beam-domain channel estimation method based on the approximate message passing algorithm, we require protection as an invention. The above are only specific implementations for specific applications, but the true spirit and scope of the present invention are not limited to this. Any person skilled in the art can modify, equivalently replace, improve, etc., to implement channel estimation methods for different applications . The present invention is defined by the claims and their equivalent technical solutions.

Claims

A deep learning beam-domain channel estimation method based on an approximate message passing algorithm is characterized by including the following steps:

(1) Construct a deep learning network structure. The deep learning network structure consists of two parts: a model-driven deep sub-network LAMP based on an approximate message passing algorithm and a data-driven deep sub-network based on residual learning ResNet; a model-driven deep sub-network The front-end output of LAMP is used as the back-end input of the data-driven deep sub-network ResNet;

(2) Model the beam domain channel according to the geometric parameters of the lens antenna, obtain the beam domain channel matrix, and obtain training data with different signal-to-noise ratios according to the MIMO system model;

(3) Use the training data with different signal-to-noise ratios described in step (2) to train the model-driven deep sub-network LAMP offline to obtain the network parameters of the model-driven deep sub-network LAMP; fix the network parameters and correct the steps (1) Perform end-to-end training on the deep learning network structure to obtain a trained deep learning network model;

(4) Use the trained deep learning network model to perform real-time beam-domain channel estimation based on the received signal at the RF chain end;

The step (1) is specifically:

(1.1) The model-driven deep sub-network LAMP is constructed based on the approximate message passing algorithm. The model-driven deep sub-network LAMP is developed by the approximate message passing algorithm. It is composed of T layers and each layer has the same structure. The t-th iteration is expressed as the t-th layer of the model-driven deep sub-network LAMP. The parameters (W, W H ) contained in the t-th iteration are replaced by the learnable parameters (β t W, Z t ) of the t-th layer, W Represents the radio frequency chain selection network of the base station, β t and Z t are the learnable parameters of the t-th layer; the beam-domain channel estimation process of the t-th layer of the model-driven deep sub-network LAMP is expressed as:

Where r is the received signal, and the subscript t represents the t-th layer of the model-driven deep sub-network LAMP,
Is the beam domain estimation channel of the t-th layer, v t is the intermediate variable, M is the number of receiving antennas, α t is the network parameter, ||·|| represents the L2 norm, ||·|| 0 represents the L0 norm; η (·;Λ) is a soft threshold function, which is defined as follows:

[η(u;λ)] j =sgn(u j )max{|u j |-λ,0}

Where u j is the j-th element of an independent variable vector of the function, λ is another independent variable of the function, which represents the threshold, and sgn(·) is a sign function;

The first-stage beam domain estimation channel output by the model-driven deep sub-network LAMP is expressed as:

among them
Represents the mapping of the model-driven deep sub-network LAMP, Θ={β t ,Z t ,α t } represents the learning parameters of the model-driven deep sub-network LAMP, and r i is the received signal;

The L2 norm is selected as the cost function, and the specific loss function is expressed as:

Where N is the number of training data, h i is a channel matrix;

(1.2) The data-driven deep sub-network ResNet is constructed based on the residual learning algorithm. The data-driven deep sub-network ResNet is composed of several residual blocks with the same structure. Each residual block has multiple convolutional layers, and each layer An activation layer is connected behind the convolutional layer. The activation function adopts tanh(·). The activation function can be guided everywhere, and the input variable is mapped to (-1,1);

The output of the data-driven deep sub-network ResNet is expressed as:

among them
In order to obtain the residual error between the estimated channel and its estimated value through ResNet,
Represents the mapping of the data-driven deep sub-network ResNet, Σ is the learning parameter of the data-driven deep sub-network ResNet, that is, the weight and bias in the convolutional layer;

(1.3) The received signal at the RF chain end passes through
with
Get the second-level beam domain estimation channel

The L2 norm is selected as the cost function, and the specific loss function is expressed as:
The method for deep learning beam-domain channel estimation based on an approximate message passing algorithm according to claim 1, wherein the step (2) is specifically:

(2.1) Model the beam domain channel according to the geometric parameters of the lens antenna, and the beam domain channel model is expressed as:

Where L represents the number of diameters, α (l) is the amplitude of the diameter l, φ (l) and θ (l) represent the incident azimuth and altitude angles of the diameter l, respectively, and N ry and N rz represent the lens antenna array along the y axis The number of antennas in the direction and the number of antennas along the z-axis, A r (φ (l) ,θ (l) ) is the response matrix of the lens antenna array, which is determined by the geometric parameters of the lens antenna. ,z) elements are expressed in the form of the product of two sinc(·) functions:

Where D Y and D Z represent the length and height of the lens antenna, and λ is the wavelength of the incident wave;

Obtain the beam domain channel matrix according to the beam domain channel model

(2.2) Obtain the received signal r at the RF chain end according to the MIMO system model,

Where s is a known pilot signal,
Is equivalent noise, n is Gaussian white noise, and h is the vectorized beam domain channel matrix;

Change the signal-to-noise ratio of the training data by adjusting the variance of the Gaussian white noise n; the vectorized beam-domain channel matrix
And the corresponding received signal vector at the RF chain end
Form the required training data
N is the number of training data.
The method for deep learning beam domain channel estimation based on approximate message passing algorithm according to claim 1, characterized in that step (3) is specifically as follows:

(3.1) Use training data with a high signal-to-noise ratio above 10dB to train the model-driven deep sub-network LAMP layer by layer. When training the t-th layer, the parameters of the first t-1 layer
Keep it unchanged, using a step-down learning rate, the initial value is a, every K times of training, the learning rate is reduced to the original τ, and the training is terminated when the normalized mean square error no longer decreases;

(3.2) Use the high signal-to-noise ratio training data described in step (3.1) to perform end-to-end training on the model-driven deep sub-network LAMP, with a learning rate of 0.1×a, and the training will stop when the normalized mean square error no longer decreases;

(3.3) Select the training data with different signal-to-noise ratios described in step (3) to optimize the deep learning network structure end-to-end. The learning rate is 0.01×a. The training is terminated when the normalized mean square error no longer decreases, and finally Get the trained deep learning network model.