CN105790813A - Method for selecting codebooks based on deep learning under large scale MIMO - Google Patents

Abstract

The invention relates to a codebook selection method based on deep learning under large-scale MIMO (Multiple-Input Multiple-Output, multiple-input multiple-output), and belongs to the technical field of wireless communication. The method includes: collecting pilot information in a test area to construct a pilot training sequence, and then obtaining a pilot training sample; performing neural network iterative learning on the pilot training sample to obtain a final network weight value; Channel, select the optimal codeword from the complete codebook. Then match the channel information between the unknown area and the test area to obtain its wireless channel, and then obtain the codeword corresponding to the wireless channel. The invention can effectively, accurately and quickly establish a wireless channel model and codebook query, avoid channel estimation in unknown areas and greatly reduce the complexity of channel selection codebooks in unknown areas.

Description

A codebook selection method based on deep learning under massive MIMO

technical field

The invention belongs to the technical field of wireless communication, and relates to a codebook selection method based on deep learning under massive MIMO (Multiple-Input Multiple-Output, multiple-input multiple-output).

Background technique

Any communication system, the channel is an essential component. The wireless channel is a typical "variable parameter channel". The characteristics of the wireless channel are closely related to the propagation environment, such as: terrain, surface features, climate characteristics, electromagnetic interference, communication object moving speed and frequency band used. The communication capability and quality of service (Quality of Service, QoS) of the wireless communication system are closely related to the performance of the wireless channel. Therefore, in order to transmit useful information with high quality and large capacity as possible on limited spectrum resources, it is necessary to have a good grasp of the characteristics of the wireless channel, especially in the era of big data, and at the same time ensure the error of the obtained wireless channel as much as possible The rate is small.

The wireless channel model is an abstract description of the wireless channel after fully understanding the wireless propagation environment and propagation characteristics, which can well reflect some important properties of the wireless propagation environment. The establishment of wireless channel model mainly depends on channel detection. At present, the existing methods for establishing wireless channel propagation models include: statistical model, deterministic model and semi-deterministic model.

However, the above-mentioned existing methods for establishing wireless channel propagation models have some disadvantages. For example, these methods are based on the theory of electromagnetic wave propagation and are analyzed under some simplified conditions to obtain a method for establishing a wireless channel model. The actual mobile propagation environment is ever-changing, which greatly limits the application range of these theoretical results. It can only be carried out for a specific environment and a single link, and the channel characteristics and directional channel characteristics in high-speed mobile scenarios are described. Not comprehensive enough. On the other hand, the existing channel model building methods need to fully explore the causal relationship between the sending and receiving ends. It collects the signals at the receiving and sending ends, and analyzes the sending and receiving signals to establish the causal relationship between the sending and receiving ends. Due to the limited samples collected and based on assumptions, the results obtained will be affected. In the era of small data and insufficient computer capabilities, most analyzes are limited to seeking simple linear relationships.

In massive MIMO systems, due to the large number of antennas, the H-dimension of the channel array increases rapidly, so the precoding technology based on non-codebook is no longer applicable, and the linear precoding technology based on codebook has become the focus of attention. At present, the commonly used methods for generating codebooks are: based on Grassmannian subspace packing, DFT, etc. However, the former generally uses exhaustive search to find the optimal codeword, such as random search, alternate prediction, and Lloyd's iterative algorithm. The computational burden of these algorithms will increase sharply with the increase in the number of transmitting antennas. While DFT systematically provides high chordal distances between precoding vectors, the average bit error rate is easily affected when transmit antennas suffer from high spatial correlation. In view of the above problems, it has become an urgent need to propose a precoding method with low computational load and anti-spatial correlation.

Contents of the invention

In view of this, the object of the present invention is to provide a codebook selection method based on deep learning under a large-scale MIMO (Multiple-InputMultiple-Output, multiple-input multiple-output), the method can effectively establish a wireless channel propagation model and a codebook Inquire.

To achieve the above object, the present invention provides the following technical solutions:

A codebook selection method based on deep learning under massive MIMO, the method comprising the following steps:

S1: Information collection step: the information collection system collects the pilot information of the user terminal in the test area;

S2: Obtain training samples: construct a pilot training sequence according to the pilot information, and then obtain pilot training samples;

S3: Initialize the neural network: initialize the parameters of the neural network model;

S4: Neural network learning: Neural network deep learning is carried out from the pilot training samples to obtain the final network weight value;

S5: Construct a complete codebook: use the improved DFT (DiscreteFourierTransform, discrete Fourier transform) method to construct a codebook suitable for all channel states;

S6: Codeword selection: according to the channel output by the neural network after learning, codeword selection is performed from the complete codebook;

S7: Establishing a correlation relationship: constructing a correlation relationship between the pilot information of the test area and the wireless channel characteristic information;

S8: Channel matching step: match the channel in the unknown area with the existing wireless channel, and then select the code word corresponding to the channel in the unknown area.

Further, in step S1, the test areas are divided into four types during information collection: suburban macro cells (suburbanmacro), urban macro cells (urbanmacro, UM-a), urban micro cells (urbanmicro, UM-i) and High-speed scene (highrisescenario).

Further, in step S3, the parameters of the initialized neural network model specifically include: learning rate η, bias value δ, weight coefficient ω _ij ∈ (0,1) of input layer i node and hidden layer j node, hidden Contains the weight coefficient ω _jl ∈(0,1) of layer j and output layer node l, where i,j,k∈N+ and ∑|ω|=M (M is a constant), the maximum number of iterations l _max , the initial error Value e=0, neuron activation function f(.) adopts threshold function, linear function or Sigmoid function.

Further, in step S4, the neural network deep learning specifically includes:

S41: The pilot training sample P is used as the input of the neural network, H=[H ₀ ,H ₁ ,...,H _N ] is the estimated target value of the neural network, is the estimated output value of the neural network;

S42: Perform in-depth training of parameters based on the error between the model output and the target value, the maximum number of iterations, and the constraints of weight values until the accuracy requirements are met;

S43: Every time it is performed, the number of iterations is increased by 1, that is, l=l+1; when the number of iterations l≤l _max or e(l)≤τ _max , the training ends, otherwise return to step S42;

S44: After steps S41, S42, and S43, obtain the weight coefficient of the target update; after the learning phase is completed, the neural network uses the pilot P in the test area to estimate and will Stored in the Shark database based on the Spark cluster, the Shark database provides users with channel information query services.

Further, in step S5, the code book suitable for all channel states constructed by the improved DFT method is as follows:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\sqrt{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}\left[\begin{array}{ccccc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; \&omega;</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 4</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]$

F = WF _DFT

Among them, W(∈M _t ×M _t ) is a unitary matrix, which satisfies U=W∑V ^H (U∈(M _t ×M _t ), whose elements obey CN(0,1)).

Further, in step S6, the code word selection includes: after the neural network learning is completed, the output value of the neural network is the channel estimated by using the pilot training samples; the code word selection is performed according to the code selection criterion, and the selected The optimal codewords are placed in the Shark database to provide users with codeword information query services.

Further, in step S7, the channel information is used to construct the correlation between the pilot information of the test area and the wireless channel characteristic information, which specifically includes: according to the pilot correlation characteristics of the test area, the pilots in the test area are divided into A plurality of representative reference pilot patterns; the channel information of the reference pilot patterns in the test area is obtained from the wireless channel model of the test area, and the channel characteristics corresponding to each reference pattern are obtained, and the channel characteristics corresponding to the reference patterns are stored In the reference channel information database.

Further, in step S8, the matching of the channel in the unknown area with the existing wireless channel, and then selecting the codeword corresponding to the channel in the unknown area specifically includes:

S81: Perform feature matching according to the pilot information of the unknown area and the reference pilot pattern in the test area;

S82: Determine whether the similarity between the pilot information of the unknown area and the reference pilot pattern feature of the test area is less than the set threshold, if less, the matching is successful; otherwise, reselect the reference pilot pattern until it meets the threshold less than the set threshold;

S83: When the pilot information in the unknown area matches the pilot pattern feature in the test area successfully, determine the channel feature corresponding to the pattern in the reference channel information database as the channel feature in the unknown area, and synthesize the channel features to obtain radio channels in the unknown zone;

S84: Obtain the optimal codeword from the Shark database storing the optimal codeword information according to the wireless channel, and feed it back to the base station (BS) in the unknown area.

The beneficial effect of the present invention is that the present invention can effectively, accurately and quickly establish a wireless channel model and codebook query, avoid channel estimation in unknown areas and greatly reduce the complexity of channel selection codebooks in unknown areas.

Description of drawings

In order to make the purpose, technical scheme and beneficial effect of the present invention clearer, the present invention provides the following drawings for illustration:

Fig. 1 is a schematic flow sheet of the method of the present invention;

Fig. 2 is the flow chart of establishing the wireless channel in the test area;

Fig. 3 is a flow chart of the steps of neural network deep learning;

FIG. 4 is a flowchart of a codebook precoding method under massive MIMO;

Fig. 5 is a flowchart of a matching model between pilot information and wireless channel characteristic information in an unknown area.

detailed description

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of the method of the present invention, as shown in the figure, the method specifically includes the following steps:

S1: Information collection step: the information collection system collects the pilot information of the user terminal in the test area;

S2: Obtain training samples: construct a pilot training sequence according to the pilot information, and then obtain pilot training samples;

S3: Initialize the neural network: initialize the parameters of the neural network model;

S4: Neural network learning: Neural network deep learning is carried out from the pilot training samples to obtain the final network weight value;

S5: Construct a complete codebook: use the improved DFT (DiscreteFourierTransform, discrete Fourier transform) method to construct a codebook suitable for all channel states;

S6: Codeword selection: according to the channel output by the neural network after learning, codeword selection is performed from the complete codebook;

S7: Establishing a correlation relationship: constructing a correlation relationship between the pilot information of the test area and the wireless channel characteristic information;

S8: Channel matching step: match the channel in the unknown area with the existing wireless channel, and then select the code word corresponding to the channel in the unknown area.

Figure 2 is a flow chart of wireless channel establishment in the test area, including:

The test areas are divided into four categories: suburban macro cells, urban macro cells, urban micro cells, and high-speed scenarios. Take UM-i as an example for information collection, and the rest are similar;

The test area is divided into four categories: suburban macro cell (suburbanmacro), urban macro cell (urbanmacro, UM-a), urban micro cell (urbanmicro, UM-i), high-speed scenario (highrisescenario), here UM-i is used as the Example for analysis, and the rest are similar;

The information collection system collects the pilot information of the user terminal in the test area;

Construct a pilot training sequence from the pilot information, and then obtain a pilot training sample;

The pilot training sample is used as the input of the neural network for deep learning of the neural network, and the output value of the neural network is obtained after the learning is completed, which is the estimated channel.

Fig. 3 is the flow chart of the steps of neural network deep learning, and described neural network deep learning specifically includes:

S41: The pilot training sample P is used as the input of the neural network, H=[H ₀ ,H ₁ ,...,H _N ] is the estimated target value of the neural network, is the estimated output value of the neural network;

S42: Carry out in-depth training of parameters based on the error between the model output and the target value, the maximum number of iterations, and the constraints of weight values until the accuracy requirements are met;

S43: Every time it is performed, the number of iterations is increased by 1, that is, l=l+1; when the number of iterations l≤l _max or e(l)≤τ _max , the training ends, otherwise return to step S42;

S44: After steps S41, S42, and S43, obtain the weight coefficient of the target update; after the learning phase is completed, the neural network uses the pilot P in the test area to estimate and will Stored in the Shark database based on the Spark cluster, the Shark database provides users with channel information query services.

Figure 4 is a flow chart of the codebook precoding method under massive MIMO. The codebook is constructed by the improved DFT method, and the codebook is placed at both ends of the transceiver:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\sqrt{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}\left[\begin{array}{ccccc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; \&omega;</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 4</span>}}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}& \mathrm{<span\; class="notranslate">\; ...</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]$

F = WF _DFT

Among them, W(∈M _t ×M _t ) is a unitary matrix, which satisfies U=WΣV ^H (U∈(M _t ×M _t ), whose elements obey CN(0,1)).

The user terminal feeds back the index of the optimal codeword output by the neural network in the codebook to the base station (BS), and stores the information of the optimal codeword in the Shark database, which provides the user with the optimal codeword information. Query service.

Fig. 5 is the flow chart of the matching model between the pilot information of the unknown area and the wireless channel characteristic information. In this embodiment, the specific process is as follows:

According to the pilot correlation characteristics of the UM-i test area, the pilots in the test area are divided into several representative reference pilot patterns;

The channel information of the reference pilot patterns in the test area is obtained from the wireless channel of the UM-i test area, the channel characteristics corresponding to each reference pattern are obtained, and the channel characteristics corresponding to the reference patterns are stored in the reference channel information database.

performing feature matching with reference pilot patterns in the UM-i test zone according to the pilot information in the unknown zone;

Judging whether the similarity between the pilot information of the unknown area and the reference pilot pattern feature of the test area is less than the set threshold, if less, the matching is successful; otherwise, reselecting the reference pilot pattern until it meets the threshold less than the set;

When the pilot information in the unknown area and the pilot pattern feature in the test area are successfully matched, the channel feature corresponding to the pattern in the reference channel information database is determined as the channel feature of the unknown area, and the channel features are synthesized to obtain the unknown wireless channels in the zone;

According to the wireless channel, in the Shark database storing the optimal code word information, the optimal code word is obtained and fed back to the BS in the unknown area.

Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that it can be described in terms of form and Various changes may be made in the details without departing from the scope of the invention defined by the claims.

Claims (8)

1. A codebook selection method based on deep learning under massive MIMO, characterized in that: the method may further comprise the steps: S1: Information collection step: the information collection system collects the pilot information of the user terminal in the test area; S2: Obtain training samples: construct a pilot training sequence according to the pilot information, and then obtain pilot training samples; S3: Initialize the neural network: initialize the parameters of the neural network model; S4: Neural network learning: Neural network deep learning is carried out from the pilot training samples to obtain the final network weight value; S5: Construct a complete codebook: use the improved DFT (DiscreteFourierTransform, discrete Fourier transform) method to construct a codebook suitable for all channel states; S6: Codeword selection: according to the channel output by the neural network after learning, codeword selection is performed from the complete codebook; S7: Establishing a correlation relationship: constructing a correlation relationship between the pilot information of the test area and the wireless channel characteristic information; S8: Channel matching step: match the channel in the unknown area with the existing wireless channel, and then select the code word corresponding to the channel in the unknown area. 2. the codebook selection method based on deep learning under a kind of massive MIMO according to claim 1 is characterized in that: in step S1, when carrying out information collection, test area is divided into four classes: suburban macro cell (suburbanmacro ), urban macro cells (urbanmacro, UM-a), urban micro cells (urbanmicro, UM-i) and high-speed scenarios (highrisescenario). 3. the codebook selection method based on deep learning under a kind of massive MIMO according to claim 2, is characterized in that: in step S3, described initialization neural network model parameter specifically comprises: learning rate n, offset value δ, the weight coefficient ω _ij ∈ (0,1) of the input layer i node and the hidden layer j node, the weight coefficient ω _jl ∈ (0,1) of the hidden layer j and the output layer node k, where i, j, k∈N ⁺ , N ⁺ is a positive integer, and Σ|ω|=M (M is a constant), the maximum number of iterations l _max , the initial error value e=0, the neuron activation function f(.) adopts the threshold function , linear function or sigmoid function. 4. the codebook selection method based on deep learning under a kind of massive MIMO according to claim 3, is characterized in that: in step S4, described neural network deep learning specifically comprises: S41: The pilot training sample P is used as the input of the neural network, and the channel array H=[H ₀ , H ₁ ,...,H _N ] is the estimated target value of the neural network, is the estimated output value of the neural network; S42: Carry out in-depth training of parameters based on the maximum error τ _max between the model output and the target value, the maximum number of iterations and weight value constraints until the accuracy requirements are met; S43: each time, the number of iterations is increased by 1, i.e. l=l+1; when the number of iterations l≤l _max or e(l)≤τ _max (τ _max maximum error), the training ends, otherwise return to step S42; S44: After steps S41, S42, and S43, obtain the weight coefficient of the target update; after the learning phase is completed, the neural network uses the pilot P in the test area to estimate and will Stored in the Shark database based on the Spark cluster, the Shark database provides users with channel information query services. 5. the codebook selection method based on deep learning under a kind of massive MIMO according to claim 4, is characterized in that: in step S5, the described codebook that is suitable for all channel states with improved DFT method construction is as follows: ${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; f</span>}\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\sqrt{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}}\left[\begin{array}{ccccc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}& <span\; class="notranslate">\; ...</span>& \mathrm{<span\; class="notranslate">\; 1</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; \&omega;</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& <span\; class="notranslate">\; ...</span>& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 4</span>}}& <span\; class="notranslate">\; ...</span>& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ <span\; class="notranslate">\; ...</span>& <span\; class="notranslate">\; ...</span>& <span\; class="notranslate">\; ...</span>& <span\; class="notranslate">\; ...</span>& <span\; class="notranslate">\; ...</span>\\ \mathrm{<span\; class="notranslate">\; 1</span>}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}& <span\; class="notranslate">\; ...</span>& {\mathrm{<span\; class="notranslate">\; \&omega;</span>}}^{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]$ F = WF _DFT Among them, F _DFT is the Fourier transform codebook, F is the modified Fourier transform codebook, M _t is the number of transmission days, W(∈M _t ×M _t ) is a unitary matrix, satisfying U=WΣV ^H (U∈(M _t ×M _t ) can be subjected to singular value decomposition, and its element u _k obeys the independent and identically distributed complex Gaussian distribution with zero mean and one variance, that is, u _k ~CN(0,1)). 6. The codebook selection method based on deep learning under a kind of massive MIMO according to claim 5, characterized in that: in step S6, the codeword selection comprises: after the neural network learning is completed, the output of the neural network The value is the channel estimated by using the pilot training samples; the code word is selected according to the code selection criterion, and the selected optimal code word is placed in the Shark database to provide users with code word information query services. 7. The codebook selection method based on deep learning under a kind of massive MIMO according to claim 6, is characterized in that: in step S7, by channel information, construct the pilot information of test area and wireless channel characteristic information Specifically, it includes: according to the pilot correlation characteristics of the test area, the pilots in the test area are divided into several representative reference pilot patterns; the reference pilot patterns in the test area are obtained by the wireless channel model of the test area. channel information of the pilot patterns, obtain the channel characteristics corresponding to each reference pattern, and store the channel characteristics corresponding to the reference patterns in the reference channel information database. 8. The codebook selection method based on deep learning under a kind of massive MIMO according to claim 7, characterized in that: in step S8, the channel in the unknown area is matched with the existing wireless channel, and then selected The codeword corresponding to the channel out of the unknown area specifically includes: S81: Perform feature matching according to the pilot information of the unknown area and the reference pilot pattern in the test area; S82: Determine whether the similarity between the pilot information of the unknown area and the reference pilot pattern feature of the test area is less than the set threshold, if less, the matching is successful; otherwise, reselect the reference pilot pattern until it meets the threshold less than the set threshold; S83: When the pilot information in the unknown area matches the pilot pattern feature in the test area successfully, determine the channel feature corresponding to the pattern in the reference channel information database as the channel feature in the unknown area, and synthesize the channel features to obtain radio channels in the unknown zone; S84: Obtain the optimal codeword from the Shark database storing the optimal codeword information according to the wireless channel, and feed it back to the base station (BS) in the unknown area.