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

Abstract

The invention relates to a method for selecting codebooks based on deep learning under large scale MIMO(Multiple-Input Multiple-Output) and belongs to the technical field of wireless communication. The method comprises following steps: acquiring pilot frequency information of a test zone to establish a pilot frequency training sequence and further obtaining a pilot frequency training sample; performing neural network iteration learning to the pilot frequency sample to obtain a final network weight value; selecting optical code words from a complete codebook according to the signal channel output by the neural network after learning. performing signal channel information matching to an unknown zone and the test zone to obtain a wireless signal channel thereof, and further obtaining code words corresponding to the wireless signal channel. By means of the method, wireless signal model and codebook query can be effectively, accurately and quickly established to avoid signal channel estimation of unknown zones and greatly reduce the complexity of unknown zone signal channel codebook selection.

Description

A kind of codebook selecting method based on degree of depth study under extensive MIMO

Technical field

The invention belongs to wireless communication technology field, relate to a kind of extensive MIMO (Multiple-Input Multiple-Output, Multiple-input and multiple-output) under based on the degree of depth study codebook selecting method.

Background technology

Any one communication system, channel is requisite ingredient.Wireless channel is typical " variable-parameter channel ", nothing The characteristic of line channel and communication environments, such as: landform, atural object, climate characteristic, electromagnetic interference situation, communication body translational speed and making Frequency range etc. closely related.The communication capacity of wireless communication system, service quality (Quality of Service, QoS) etc. are all Closely related with the quality of radio channel performance.Therefore, want high-quality, Large Copacity as far as possible on limited frequency spectrum resource to pass Defeated useful information, it is necessary to grasp the characteristic of wireless channel well, especially at big data age, to ensure the most as far as possible The error rate of the wireless channel obtained is less.

Wireless channel model be wireless propagation environment and propagation characteristic had be fully understood by after, one of wireless channel abstract is retouched State, some critical natures of wireless propagation environment can be reflected well.The foundation of wireless channel model depends on channel detection. At present, the existing method setting up wireless channel propagation model has: statistical model, deterministic models and semidefiniteness model.

But the above-mentioned existing method setting up wireless channel propagation model there are disadvantages that, as these methods are based on electromagnetic wave propagation Theory, analyzes the method for building up drawing wireless channel model under some reduced conditions.And actual mobile circumstances is thousand changes ten thousand Change, limit the range of application of these notional results significantly, can only carry out for certain specific environment, single link, Describe the characteristic of channel under high-speed mobile scene, the directivity characteristic of channel is not comprehensively accurate.On the other hand, existing letter The method for building up of road model needs fully to excavate the cause effect relation of sending and receiving end.It analyzes sending and receiving letter by gathering the signal of sending and receiving end Number set up the cause effect relation of transmitting-receiving two-end.Because of gather Finite Samples, and based on the assumption that condition so as to get result can be impacted. In the small data epoch, computer capacity is not enough, and major part is analyzed and is only limitted to seek simple linear relationship.

In extensive mimo system, make because number of antennas is huge channel battle array H dimension become rapidly big, based on non-code book prelist Code technology is the most applicable, and Linear Precoding based on code book has become focus of attention.The most conventional side producing code book Method has: based on Grassmannian subspace packing, DFT etc..But the former looks for optimum code word at general exhaustive search, Such as random search, alternately prediction, Selwyn Lloyd iterative algorithm, the computation burden of these algorithms will increase urgency along with launch antenna number Increase severely big.And DFT provides high chordal distance by the mode of system between precoding vectors, but average error rate is easily being sent out Penetrate antenna to suffer to receive impact during high spatial dependency.For problem above, a kind of low amount of calculation, anti-spatial coherence are proposed Method for precoding has become urgent needs.

Summary of the invention

In view of this, it is an object of the invention to provide a kind of extensive MIMO (Multiple-Input Multiple-Output, Multiple-input and multiple-output) under based on the degree of depth study codebook selecting method, the method can set up wireless channel propagation model effectively And code book inquiry.

For reaching above-mentioned purpose, the present invention provides following technical scheme:

A kind of codebook selecting method based on degree of depth study under extensive MIMO, the method comprises the following steps:

S1: information gathering step: by the pilot frequency information of user side in information acquisition system collecting test district；

S2: obtain training sample: build pilot training sequence according to pilot frequency information, and then obtain pilot tone training sample；

S3: initialize neutral net: initialize neural network model parameter；

S4: neural network learning: carried out neutral net degree of depth study by pilot tone training sample, obtains final network weight weight values；

S5: construct complete code book: by DFT (Discrete Fourier Transform, the discrete Fourier transform) method improved Structure is suitable for the code book of all channel status；

S6: codeword selection: according to the channel of the neutral net output after study, carry out codeword selection from complete code book；

The foundation of S7: dependency relation: build the dependency relation between the pilot frequency information of described test section and characteristics of radio channels information；

S8: channel matched step: the channel of unknown area is mated with existing wireless channel, and then select the channel pair of unknown area The code word answered.

Further, in step sl, when carrying out information gathering, test section is divided into four classes: suburb macrocell (suburban Macro), urban macro community (urban macro, UM-a), urban microcell (urban micro, UM-i) and High-speed Circumstance (high rise scenario)。

Further, in step s3, described initialization neural network model parameter specifically includes: learning rate η, bias δ, Input layer i-node and the weights coefficient ω of hidden layer j node_ij∈ (0,1), hidden layer j and the weights coefficient of output layer node l ω_jl∈ (0,1), wherein i, j, k ∈ N+ and ∑ | ω |=M (M is constant), maximum iteration time l_max, Initial value e=0, god Threshold function table, linear function or Sigmoid function is used through unit's activation primitive f (.).

Further, in step s 4, the study of the described neutral net degree of depth specifically includes:

S41: pilot tone training sample P is as the input of neutral net, H=[H₀,H₁,...,H_N] it is the estimation target of neutral net Value,Estimation output valve for neutral net；

S42: exported by model and error, maximum iteration time and weighted value constraints between desired value carries out the degree of depth instruction of parameter Practice, until being met required precision；

S43: often carry out once, iterations adds 1 i.e. l=l+1；As iterations l≤l_maxOr e (l)≤τ_maxTime terminate training, Otherwise return step S42；

S44: obtain the weights coefficient of target update after step S41, S42, S43；After the study stage completes, neutral net profit Estimate by pilot tone P of test sectionAnd willIt is stored in Shark data base based on Spark cluster, Shark data base Provide the user the inquiry service of channel information.

Further, in step s 5, the code book of the applicable all channel status of DFT method structure of described improvement is as follows:

$F_{DFT}=1/\sqrt{M_t}\left[\begin{array}{ccccc}1& 1& 1& \mathrm{...}& 1\\ 1& \mathrm{\ω}& {\mathrm{\ω}}^2& \mathrm{...}& {\mathrm{\ω}}^{M_t-1}\\ 1& {\mathrm{\ω}}^2& {\mathrm{\ω}}^4& \mathrm{...}& {\mathrm{\ω}}^{2\left(M_t-1\right)}\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ 1& {\mathrm{\ω}}^{M_t-1}& {\mathrm{\ω}}^{2(M_t-1)}& \mathrm{...}& {\mathrm{\ω}}^{\left(M_t-1\right)\left(M_t-1\right)}\end{array}\right]$

F=WF_DFT

Wherein, W (∈ M_t×M_t) it is unitary matrice, meet U=W ∑ V^H(U∈(M_t×M_t), its element obeys CN (0,1)).

Further, in step s 6, described codeword selection includes: after neural network learning completes, and the output valve of neutral net is i.e. For the channel utilizing pilot tone training sample to estimate；Carry out codeword selection according to code selection criterion, and the optimum code word selected is placed on In Shark data base, provide the user the inquiry service of codeword information.

Further, in the step s 7, by channel information, the phase between the pilot frequency information of test section with characteristics of radio channels information is built Pass relation, specifically includes: according to the pilot correlation feature of test section, the pilot tone in test section be divided into multiple representative Reference pilot pattern；Obtained the channel information of this test section internal reference pilot frequency design by the wireless channel model of test section, obtain The channel characteristics that each reference pattern is corresponding, and channel characteristics corresponding for reference pattern is stored in reference channel information data base.

Further, in step s 8, described the channel of unknown area is mated with existing wireless channel, and then select unknown area The code word that channel is corresponding specifically includes:

S81: carry out characteristic matching with the reference pilot pattern in test section according to the pilot frequency information of described zone of ignorance；

S82: judge that whether the similarity between the pilot frequency information of zone of ignorance and the reference pilot pattern characteristics of test section is less than setting Threshold value, if less than, the match is successful；Otherwise, again choose reference pilot pattern, until meeting less than the threshold value set；

S83: when after the pilot frequency information in unknown area and the success of the pilot frequency design characteristic matching in test section, by reference channel information number It is defined as the channel characteristics of unknown area according to the channel characteristics that this pattern in storehouse is corresponding, carries out channel characteristics comprehensively, obtaining this unknown Wireless channel in district；

S84: according to this wireless channel, in the Shark data base storing optimum codeword information, obtain optimum code word, and by it Feed back to the base station (BS) in this unknown area.

The beneficial effects of the present invention is: the present invention can set up wireless channel model effectively, accurately and rapidly and inquire about with code book, The channel avoiding unknown area is estimated and greatly reduces the complexity of unknown area Channel assignment code book.

Accompanying drawing explanation

In order to make the purpose of the present invention, technical scheme and beneficial effect clearer, the present invention provides drawings described below to illustrate:

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

Fig. 2 is the wireless channel Establishing process figure of test section；

Fig. 3 is the flow chart of steps of neutral net degree of depth study；

Fig. 4 is codebook precoding method flow diagram under extensive MIMO；

Fig. 5 is the Matching Model flow chart between unknown area pilot frequency information and characteristics of radio channels information.

Detailed description of the invention

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

Fig. 1 is the schematic flow sheet of the method for the invention, as it can be seen, this method specifically includes following steps:

S1: information gathering step: by the pilot frequency information of user side in information acquisition system collecting test district；

S2: obtain training sample: build pilot training sequence according to pilot frequency information, and then obtain pilot tone training sample；

S3: initialize neutral net: initialize neural network model parameter；

S4: neural network learning: carried out neutral net degree of depth study by pilot tone training sample, obtains final network weight weight values；

S5: construct complete code book: by DFT (Discrete Fourier Transform, the discrete Fourier transform) method improved Structure is suitable for the code book of all channel status；

S6: codeword selection: according to the channel of the neutral net output after study, carry out codeword selection from complete code book；

The foundation of S7: dependency relation: build the dependency relation between the pilot frequency information of described test section and characteristics of radio channels information；

S8: channel matched step: the channel of unknown area is mated with existing wireless channel, and then select the channel pair of unknown area The code word answered.

Fig. 2 is the wireless channel Establishing process figure of test section, including:

Test section is divided into four classes: suburb macrocell, urban macro community, urban microcell, High-speed Circumstance, as a example by UM-i Carrying out information gathering, remaining is similar；

Test section is divided into four classes: suburb macrocell (suburban macro), urban macro community (urban macro, UM-a), city Microcell, district (urban micro, UM-i), High-speed Circumstance (high rise scenario), be analyzed here as a example by UM-i, its Remaining is similar；

The pilot frequency information of user side in information acquisition system collecting test district；

Built pilot training sequence by pilot frequency information, and then obtain pilot tone training sample；

Pilot tone training sample, as the input of neutral net, carries out neutral net degree of depth study, and study obtains neutral net after terminating Output valve, is the channel estimated.

Fig. 3 is the flow chart of steps of neutral net degree of depth study, and the study of the described neutral net degree of depth specifically includes:

S41: pilot tone training sample P is as the input of neutral net, H=[H₀,H₁,...,H_N] it is the estimation target of neutral net Value,Estimation output valve for neutral net；

S42: exported by model and error, maximum iteration time and weighted value constraints between desired value carries out the degree of depth instruction of parameter Practice, until being met required precision；

S43: often carry out once, iterations adds 1 i.e. l=l+1；As iterations l≤l_maxOr e (l)≤τ_maxTime terminate training, Otherwise return step S42；

S44: obtain the weights coefficient of target update after step S41, S42, S43；After the study stage completes, neutral net profit Estimate by pilot tone P of test sectionAnd willIt is stored in Shark data base based on Spark cluster, Shark data base Provide the user the inquiry service of channel information.

Fig. 4 is codebook precoding method flow diagram under extensive MIMO, by the DFT method structure code book improved, and by code Originally transmitting-receiving two-end it is placed on:

$F_{DFT}=1/\sqrt{M_t}\left[\begin{array}{ccccc}1& 1& 1& \mathrm{...}& 1\\ 1& \mathrm{\ω}& {\mathrm{\ω}}^2& \mathrm{...}& {\mathrm{\ω}}^{M_t-1}\\ 1& {\mathrm{\ω}}^2& {\mathrm{\ω}}^4& \mathrm{...}& {\mathrm{\ω}}^{2\left(M_t-1\right)}\\ \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}& \mathrm{...}\\ 1& {\mathrm{\ω}}^{M_t-1}& {\mathrm{\ω}}^{2(M_t-1)}& \mathrm{...}& {\mathrm{\ω}}^{\left(M_t-1\right)\left(M_t-1\right)}\end{array}\right]$

F=WF_DFT

Wherein, W (∈ M_t×M_t) it is unitary matrice, meet U=W Σ V^H(U∈(M_t×M_t), its element obeys CN (0,1)).

The optimum code word that neutral net is exported by user side index in the codebook feeds back to base station end (BS), and by optimum code word Information stores in Shark data base, and this Shark data base provides the user the inquiry service of optimum code word.

Fig. 5 is the Matching Model flow chart between unknown area pilot frequency information and characteristics of radio channels information, in the present embodiment, specifically Flow process is as follows:

Pilot correlation feature according to UM-i test section, is divided into multiple representative reference to lead by the pilot tone in test section Frequently pattern；

Obtained the channel information of this test section internal reference pilot frequency design by the wireless channel of UM-i test section, obtain each with reference to figure The channel characteristics that case is corresponding, and channel characteristics corresponding for reference pattern is stored in reference channel information data base.

Pilot frequency information according to described unknown area carries out characteristic matching with the reference pilot pattern in UM-i test section；

Judge that whether the similarity between the pilot frequency information of zone of ignorance and the reference pilot pattern characteristics of test section is less than the threshold set Value, if less than, the match is successful；Otherwise, again choose reference pilot pattern, until meeting less than the threshold value set；

When after the pilot frequency information in unknown area and the success of the pilot frequency design characteristic matching in test section, by reference channel information data base In channel characteristics corresponding to this pattern be defined as the channel characteristics of unknown area, channel characteristics is carried out comprehensively, obtains in this unknown area Wireless channel；

According to this wireless channel, in the Shark data base storing optimum codeword information, obtain optimum code word, and fed back To the BS in this unknown area.

Finally illustrating, preferred embodiment above is only in order to illustrate technical scheme and unrestricted, although by above-mentioned The present invention is described in detail by preferred embodiment, it is to be understood by those skilled in the art that can in form and In details, it is made various change, without departing from claims of the present invention limited range.

Claims (8)

1. codebook selecting method based on degree of depth study under an extensive MIMO, it is characterised in that: the method includes following Step:

S1: information gathering step: by the pilot frequency information of user side in information acquisition system collecting test district；

S2: obtain training sample: build pilot training sequence according to pilot frequency information, and then obtain pilot tone training sample；

S3: initialize neutral net: initialize neural network model parameter；

S4: neural network learning: carried out neutral net degree of depth study by pilot tone training sample, obtains final network weight weight values；

S5: construct complete code book: by DFT (Discrete FourierTransform, the discrete Fourier transform) method improved Structure is suitable for the code book of all channel status；

S6: codeword selection: according to the channel of the neutral net output after study, carry out codeword selection from complete code book；

The foundation of S7: dependency relation: build the dependency relation between the pilot frequency information of described test section and characteristics of radio channels information；

S8: channel matched step: the channel of unknown area is mated with existing wireless channel, and then select the channel pair of unknown area The code word answered.

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 1 It is: in step sl, when carrying out information gathering, test section is divided into four classes: suburb macrocell (suburbanmacro), city District's macrocell (urbanmacro, UM-a), urban microcell (urbanmicro, UM-i) and High-speed Circumstance (highrise scenario).

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 2 Being: in step s3, described initialization neural network model parameter specifically includes: learning rate η, bias δ, input layer I-node and the weights coefficient ω of hidden layer j node_ij∈ (0,1), hidden layer j and the weights coefficient ω of output layer node k_jl∈ (0,1), Wherein i, j, k ∈ N⁺, N⁺For positive integer, and Σ | ω |=M (M is constant), maximum iteration time l_max, Initial value e=0, Neuron activation functions f (.) uses threshold function table, linear function or Sigmoid function.

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 3 Being: in step s 4, the study of the described neutral net degree of depth specifically includes:

S41: pilot tone training sample P as the input of neutral net, channel battle array H=[H₀,H₁,...,H_N] it is the estimation of neutral net Desired value,Estimation output valve for neutral net；

S42: exported by model and maximum error τ between desired value_max, maximum iteration time and weighted value constraints carry out parameter Degree of depth training, until be met required precision；

S43: often carry out once, iterations adds 1 i.e. l=l+1；As iterations l≤l_maxOr e (l)≤τ_max(τ_maxMaximum is by mistake Difference) time terminate training, otherwise return step S42；

S44: obtain the weights coefficient of target update after step S41, S42, S43；After the study stage completes, neutral net profit Estimate by pilot tone P of test sectionAnd willIt is stored in Shark data base based on Spark cluster, Shark data base Provide the user the inquiry service of channel information.

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 4 Being: in step s 5, the code book that the DFT method structure of described improvement is suitable for all channel status is as follows:

$F_{DFT}=1/\sqrt{M_t}\left[\begin{array}{ccccc}1& 1& 1& ...& 1\\ 1& \mathrm{\ω}& {\mathrm{\ω}}^2& ...& {\mathrm{\ω}}^{M_t-1}\\ 1& {\mathrm{\ω}}^2& {\mathrm{\ω}}^4& ...& {\mathrm{\ω}}^{2(M_t-1)}\\ ...& ...& ...& ...& ...\\ 1& {\mathrm{\ω}}^{M_t-1}& {\mathrm{\ω}}^{2(M_t-1)}& ...& {\mathrm{\ω}}^{(M_t-1)(M_t-1)}\end{array}\right]$

F=WF_DFT

Wherein, F_DFTFor Fourier transform code book, F is amended Fourier transform code book, M_tFor launching natural law, W(∈M_t×M_t) it is unitary matrice, meet U=W Σ V^H(U∈(M_t×M_t) singular value decomposition, its element u can be carried out_kObey 0 average, the independent identically distributed multiple Gauss distribution, i.e. u of 1 variance_k～CN (0,1)).

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 5 Being: in step s 6, described codeword selection includes: after neural network learning completes, and the output valve of neutral net is utilization The channel that pilot tone training sample estimates；Carry out codeword selection according to code selection criterion, and the optimum code word selected is placed on Shark In data base, provide the user the inquiry service of codeword information.

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 6 It is: in the step s 7, by channel information, builds the dependency relation between the pilot frequency information of test section and characteristics of radio channels information, Specifically include: according to the pilot correlation feature of test section, the pilot tone in test section is divided into multiple representative reference lead Frequently pattern；Obtained the channel information of this test section internal reference pilot frequency design by the wireless channel model of test section, obtain each reference The channel characteristics that pattern is corresponding, and channel characteristics corresponding for reference pattern is stored in reference channel information data base.

Codebook selecting method based on degree of depth study, its feature under a kind of extensive MIMO the most according to claim 7 It is: in step s 8, described the channel of unknown area is mated with existing wireless channel, and then select the channel pair of unknown area The code word answered specifically includes:

S81: carry out characteristic matching with the reference pilot pattern in test section according to the pilot frequency information of described zone of ignorance；

S82: judge that whether the similarity between the pilot frequency information of zone of ignorance and the reference pilot pattern characteristics of test section is less than setting Threshold value, if less than, the match is successful；Otherwise, again choose reference pilot pattern, until meeting less than the threshold value set；

S83: when after the pilot frequency information in unknown area and the success of the pilot frequency design characteristic matching in test section, by reference channel information number It is defined as the channel characteristics of unknown area according to the channel characteristics that this pattern in storehouse is corresponding, carries out channel characteristics comprehensively, obtaining this unknown Wireless channel in district；

S84: according to this wireless channel, in the Shark data base storing optimum codeword information, obtain optimum code word, and by it Feed back to the base station (BS) in this unknown area.