CN106899338A - User packet method based on density in extensive mimo system downlink - Google Patents

Abstract

The present invention is based on density user packet method in disclosing a kind of extensive MIMO downlinks, comprises the following steps：Step 1, all users distance each other is calculated, obtain distance matrix；Step 2, obtained according to distance matrix k dist figure, two important parameters of DBSCAN algorithms are obtained according to the user density distributed intelligence included in figure：Radius and minimum number of users threshold value；Step 3, cluster just is proceeded by from any one user later according to described two parameters, obtain final user grouping result.Using technical scheme, by finding suitable user between apart from metric form, using density clustering, all users in a cell are divided into several groups, can greatly improve system velocity by the Density Distribution information using user.

Description

User packet method based on density in extensive mimo system downlink

Technical field

Association area the invention belongs to improve channel capacity technique study in extensive MIMO, more particularly to a kind of FDD In system, density user packet method is based in extensive MIMO downlinks.

Background technology

In in the past few decades, the demand in wireless network to message transmission rate is sharply increased, but limited frequency spectrum money Source significantly limit the growth of transmission rate, and then extensive MIMO relies on its high spectrum utilization and high-energy utilization rate, It is considered as the core technology of Next-Generation Wireless Communication Systems.Extensive mimo system is equipped with by base station end and user terminal Tens up to a hundred antennas meet growing transmission rate requirements, while do not increase extra communication bandwidth.And advise greatly Mould MIMO can reduce system delay, simplify system MAC layer, and robustness is stronger.But with TDD system in it is extensive MIMO is different, for FDD system in extensive MIMO, have one and know critical defect.The up-downgoing channel of TDD system is present Interact, it is possible to use uplink carries out down channel estimation；And the up-downgoing channel of FDD system does not exist such phase Interaction, and large-scale antenna system may require that great channel estimation expense, have a strong impact on the reality of extensive MIMO Using.A joint space diversity and multiplexing are proposed in order to solve the scholars such as this problem, Ansuman Adhikary (JSDM) scheme, after be grouped for intra-cell users by the program, then using two grades of pre-coding schemes, be so as to greatly reduce The expense of the channel estimation of system.In JSDM schemes, user grouping is primarily to carry out dimension-reduction treatment to system, and two grades pre- Coding then greatly simplify the computation complexity of precoding, and in two grades of precodings, the switched-beam of the first order is encoded for disappearing Except user group between disturb, and the precoding of the second level is to eliminate the interference in group between user；But in JSDM, user The result of packet will have large effect to system velocity.

The content of the invention

The present invention proposes user of the joint based on density based on user grouping in JSDM schemes and two grades of precoding techniques The method of packet, by finding suitable user between apart from metric form, using density clustering, using the density of user All users in one cell are divided into several groups by distributed intelligence, can greatly improve system velocity.

To achieve the above object, the present invention is adopted the following technical scheme that：

Comprised the following steps based on density user packet method in a kind of extensive MIMO downlinks：

Step 1, all users distance each other is calculated, obtain distance matrix；

Step 2, obtained according to distance matrix k-dist figure, obtained according to the user density distributed intelligence included in figure Two important parameters of DBSCAN algorithms：Radius and minimum number of users threshold value；

Step 3, cluster just is proceeded by from any one user later according to described two parameters, obtain final user Group result.

Preferably, step 1 is specially：The distance between any two user is calculated, the distance matrix of K × K dimensions is obtained Dis, the distance between (k, j) individual element representation user k and user j, is expressed as [dis] in matrix_k,j=S (U_k,U_j), its In, S (U_k,U_j) according to distance between the user that is calculated of Similarity Measures based on subspace projection.

Preferably, step 2 is specially：If it is user j and k-th his nearest user that the k-dist of user j is represented The distance between, after the k-dist values for obtaining each user, by the arrangement of these value ascending orders, k-dist figures are then drawn as, obtain K-dist figures contain the Density Distribution information of user, first point in the k-dist figures in first trough is threshold value Point；Wherein, the k values in k-dist figures are equal to Minpts, and the value of the corresponding y-axis of the threshold point is exactly the eps values for obtaining.

Preferably, step 3 is specially：First randomly choose a user k as initial user, and travel through it is all not Labeled as the user of " processed ", finding out can be adjacent from the reachable user's set of the direct density of user k, the eps- for obtaining user k Domain N_eps(k), if | N_eps(k) |=1, then user k is exactly noise user, mark the user as " processed ", Ran Houzai Never it is labeled as randomly choosing next user in the user of " processed "；If 1 ＜ | N_eps(k) | ＜ Minpts, then user K is exactly edge customer, and any operation is not carried out to user k, but never labeled as under random selection in the user of " processed " One user；If | N_eps(k) | >=Minpts, then user k is exactly core customer, it is possible to start a cluster for group； A group, N are clustered since core customer k_epsK () is exactly a part for the group, then again from N_epsIn all users in (k) Core customer is looked for, if there is core customer q, then by the eps- neighborhoods N of the core customer_epsQ all users in () are also grouped into In the group, these users are collectively labeled as " processed ", and core customer is searched for from newly user is added to, repeat above-mentioned step Suddenly, constantly cluster, untill without new core customer, all users for adding the groups just constitute a user's group；One After individual group cluster terminates, then never continue above-mentioned all of step labeled as a user is randomly choosed in the user of " processed " Suddenly, until all users are marked as " processed ", at this time all users or noise user, or divided Group, so far, the user grouping stage just finishes；

After user grouping terminates, each group is assisted by asking the covariance of all users in group to be worth to each center organized Variance, i.e.,Then obtained between switched-beam matrix B, elimination group after interference using approximate BD algorithms, profit With MAX user's selection algorithm, user's selection is carried out respectively to each group, finally obtain pre-coding matrix with ZF precoding algorithms P；

After user's selection terminates, it is assumed that have S in g groups_gIndividual user is selected can to transmit data, thenTable Show the quantity of all users being selected in whole cell, useRepresent s-th instant Signal to Interference plus Noise Ratio of user in g groups (SINR), it is defined as：

Wherein P represents the signal power that base station provides,Represent user g_sPrecoding vector, be pre-coding matrix P_g's S is arranged,WithInterference and the interior interference of group between the group of user are represented respectively, are respectively defined as

Then user g_sSpeedCan be calculated by following formula respectively with speed C with system

This method is grouped using the Density Distribution information of user itself to user, and the group result for obtaining is fully demonstrated User distribution situation in the cell, good basis is provided to follow-up treatment.Test result indicate that, this method is one Determine to improve throughput of system in degree.

Brief description of the drawings

Fig. 1, monocyclic illustraton of model；

Fig. 2,6-dist scheme；

Fig. 3, final system conjunction rate compare figure；

Flow chart based on density user packet method in Fig. 4, extensive MIMO downlinks.

Specific embodiment

The embodiment of the present invention is based on density user packet method in providing a kind of extensive MIMO downlinks, using k- Dist figures estimate the Density Distribution situation of user, and user is grouped using DBSCAN algorithms then, greatly reduce big rule The computation complexity of mould mimo system.

The inventive method is the method that the user grouping based on density is proposed on the basis of JSDM schemes, JSDM schemes It is the scheme of the extensive MIMO complexities of reduction that Ansuman Adhikary et al. are proposed on the basis of channel relevancy, mainly Feature has 2：First is that all users are grouped according to the similarity of channel relevancy, and second by pre-coding matrix It is divided into two-stage, is disturbed between elimination group respectively and disturbed in organizing.The physical model of this method is in single cell FDD system, it is considered to Downlink in extensive mimo system, transmitting terminal is base station, and receiving terminal is K user, and base station end is equipped with M root antennas Even linear array, user is single-antenna subscriber, and channel model can be expressed as

Wherein z:It is a vector for K × 1 dimension, represents additive white Gaussian noise, x is the vector of M × 1 dimension, table Show the transmission signal of base station, y is the vector of K × 1 dimension, represent the signal that user terminal is received, and h_kWhat is represented is the letter of user k Road information, is a vector for M × 1 dimension.It is the Linear precoding matrix V of M × S to use dimension in base station end, therefore transmission letter X=Vd number can be expressed as, wherein d represents the data vector of S × 1 dimension, and S is the quantity of downstream data flow.Prelisted using two grades After code, pre-coding matrix can be expressed as：V=BP, wherein B are the switched-beam codings of the first order, and dimension is M × b, and P is The conventional multi-user pre-coding matrix of the second level, dimension is b × S.Then channel model can be expressed as again：

Y=[h₁,h₂...,h_K]^HBPd+z

Consider the correlation of transmission antenna, JSDM schemes employ monocyclic model, and user is relative to the azimuth of base station θ_k, angle spread is Δ_k, represent that the distance between carrier wavelength, two antennas can be expressed as λ D with λ.Then in even linear array In, the m rows of the channel covariance matrices of user k, the element of the n-th row can be calculated according to the following formula

Consider non-line-of-sight propagation model, the channel information h of user k_k:Wherein R_kIt is the channel covariancc of user k Matrix, dimension is M × M.In order to calculate covariance matrix R_k, it is considered to the monocyclic model of Fig. 1, the azimuth in user k locations is θ_k, angle spread is Δ_k, and in even linear array, the distance between two antennas is λ D, wherein λ is wavelength.According to these information, use Correlation between the m roots of family k and n-th antenna can be calculated by following formula

By R_kCarrying out Eigenvalues Decomposition can obtainWherein Λ_kIt is the diagonal matrix of r × r dimensions, its is diagonal Element is R_kNonzero eigenvalue, U_kIt is the unitary matrice high of M × r dimensions, its each row is all Λ_kCorresponding characteristic vector, and r is R_kOrder.The situation of non-line-of-sight propagation is not considered, and the channel vector of user k can be expressed as h_k:CN(0,R_k), application Karhunen-Loeve is converted, and the channel vector of user k can be expressed as

Wherein

In JSDM schemes, all of user is divided into G group in cell, and the user in same group shares a channel Covariance matrix, wherein the channel covariancc of all users in g groups is all denoted asOrder is denoted as r_g.Assuming that g There is K in group_gIndividual user, andThe switched-beam matrix of so system can be expressed as B=[B₁,...,B_G], its Middle B_gIt is M × b_gThe switched-beam matrix of the g groups of dimension, B_gIt is calculated according to second-order channel statistics information, for pressing down (INGI) is disturbed between processed group.The pre-coding matrix of simultaneity factor can be expressed as P=[P₁,...,P_G], wherein P_gIt is b_g×S_gDimension G groups pre-coding matrix, for organizing the interference (INAI) between interior user always.Herein, b_gIt is the number of wave beam in g groups Amount, S_gIt is the quantity of independent data stream in g groups, it is necessary to meetWithAnd due in a group, counting According to stream transmitted on wave beam, the independent data stream quantity of transmission must also than that can not possibly accomplish more than numbers of beams B >=S and b must be met_g≥S_g.The channel matrix of g groups is denoted asThe channel matrix of system is denoted as H= [H₁,...,H_G], the signal that then g groups are received can be expressed as：

In the method, switched-beam matrix B_gIt is calculated with diagonal (DB) algorithm of approximate block, pre-coding matrix P_gWith ZF precoding algorithms are calculated.

In JSDM schemes, all of user is assigned in different groups according to respective channel covariancc.After having divided group, The center covariance of each user's group is worth to by calculating the average of the covariance of all users in the group.User grouping exists Effect in JSDM schemes is particularly significant, and user grouping can influence switched-beam matrix B_gCalculating, can also influence user to select The result selected.The user packet method of several low complex degrees is had been proposed that now, such as K- averages (K-Means) are used Family grouping algorithm and K- central points (K-Medoids) user grouping algorithm, this method are used based on density (DBSCAN) User grouping algorithm.All users are divided into several different groups by DBSAN algorithms according to the Density Distribution of user, are calculated with first two Method compares, and the latter does not need the quantity of designated packet, and can be to find packet count automatically according to the Density Distribution of user.And Can only find that circular or spherical cluster is different from K-Means algorithms and K-Medoids, DBSCAN algorithms can be found that arbitrary shape Cluster.Meanwhile, DBSCAN algorithms can also detect that noise spot, at least 2 advantages of this characteristic：First, noise spot can shadow The accuracy in computation of Xiang Zu centers covariance, detect noise spot and give up can Shi Zu centers covariance it is more representative, from And reduce interference；Secondly, the noise spot for detecting must be rejected in user's choice phase, in advance detect noise spot And give up, unnecessary amount of calculation can be in advance reduced, computation complexity is reduced to a certain extent.

Before formally starting user grouping, it is thus necessary to determine that apart from measure between user.This method is used based on son The method for measuring similarity of space projection.Use U_kRepresent the covariance feature matrix of user k, V_gRepresent the center covariance of g groups Eigenmatrix, then the distance between user k and g groups are defined as follows：

Similar, the distance between user k and user j is defined as follows：

As can be seen that the similarity between two users is higher from definition, distance is nearer, that is, S (U_k,U_j) get over It is small.The distance matrix dis of user is can be obtained by according to the method for measuring similarity based on subspace projection.

It is grouped using DBSCAN algorithms, it is necessary to be specified 2 parameters：The eps of user's radius is represented, and represents given User minimum neighborhood number of users Minpts as kernel object in radius.In addition, there is several definition in DBSCAN algorithms Need explanation.

Define 1：The set of all users is represented with u, then use N_epsK () represents the eps- neighborhoods of user k, i.e., in user k Eps radiuses in all users set, be defined as:N_eps(k)=and j ∈ u | dis (k, j)≤eps }

Define 2：When meeting condition | N_eps(k) | during >=Minpts, user is core customer；When meeting the ＜ of condition 1 | N_eps (k) | during ＜ Minpts, user is edge customer；When meeting condition | N_eps(k) | when=1, user is noise user.Should be noted , with the carrying out of packet, edge customer may turn into a member of certain group, or as noise user.

Define 3：When meeting condition j ∈ N_eps(k) and | N_eps(k) | during >=Minpts, user j can be directly close from user k Degree is reachable.

Define 4：If there is user's chain o₁,...,o_n, wherein o₁=k, o_n=j, and for any 1 ＜ i≤n, user o_i Can be from user o_i-1Direct density is reachable, then, user j can be reachable from user's k density.

Define 5：If there is user o, user j and user k can be reachable from user's o density, then, user j and user k It is exactly what density was connected.

After specify that the several definition of the above, it is possible to start to carry out user the packet based on density, as shown in Figure 4.

The first step based on Density fraction, seeks to calculate the distance between any two user, obtain K × K dimensions away from From matrix dis, the distance between (k, j) individual element representation user k and user j, is expressed as [dis] in matrix_k,j=S (U_k, U_j), wherein S (U_k,U_j) it is exactly distance between the user being calculated according to the Similarity Measures based on subspace projection.

The second step of user grouping will determine parameter eps and Minpts.After obtaining distance matrix, parameter Minpts according to Experience is worth to, and obtaining parameter Minpts needs to scheme using k-dist.That the k-dist of user j is represented is user j with from him The distance between k-th nearest user, after obtaining the k-dist values of each user, then the arrangement of these value ascending orders is drawn as Figure, as illustrated in figure 2 of the appended drawings.The k-dist figures for obtaining contain the Density Distribution information of user.Distance in k-dist figures Suddenly first trough in the drastically change, therefore figure that indicate density is increased, first point in trough is threshold point, such as Point pointed by arrow in Fig. 2, parameter eps can be obtained according to threshold point.In the method, k values in k-dist figures etc. In Minpts, after obtaining threshold point, the value of the corresponding y-axis of threshold point is exactly the eps values for obtaining.

After determining parameter eps and Minpts, user grouping just can formally start.A user k is randomly choosed first As initial user, and travel through it is all it is unmarked be the user of " processed ", finding out can be reachable from the direct density of user k User gathers, and obtains the eps- neighborhoods N of user k_eps(k).If | N_eps(k) |=1, then user k is exactly noise user, by this User's mark is " processed ", is then never labeled as randomly choosing next user in the user of " processed " again；If 1 ＜ |N_eps(k) | ＜ Minpts, then user k is exactly edge customer, any operation is not carried out to user k, but be never labeled as Next user is randomly choosed in the user of " processed "；If | N_eps(k) | >=Minpts, then user k is exactly that core is used Family, it is possible to start a cluster for group.A group, N are clustered since core customer k_epsK () is exactly a part for the group, Then again from N_epsCore customer is looked in all users in (k), if there is core customer q, then by the eps- of the core customer Neighborhood N_epsQ all users in () are also grouped into the group, these users are collectively labeled as " processed ", and are added to from newly Core customer is searched in user, is repeated the above steps, constantly cluster, untill without new core customer, all additions should The user of group just constitutes a user's group.After one group cluster terminates, then never labeled as random in the user of " processed " One user of selection, continues above-mentioned all steps, until all users are marked as " processed ", at this time all users Or noise user, or be divided into group, so far, the user grouping stage just finishes.

After user grouping terminates, each group is assisted by asking the covariance of all users in group to be worth to each center organized Variance, i.e.,Then obtained between switched-beam matrix B, elimination group after interference using approximate BD algorithms, profit With MAX user's selection algorithm, user's selection is carried out respectively to each group, finally obtain pre-coding matrix with ZF precoding algorithms P。

After user's selection terminates, it is assumed that have S in g groups_gIndividual user is selected can to transmit data, thenTable Show the quantity of all users being selected in whole cell.WithRepresent s-th instant Signal to Interference plus Noise Ratio of user in g groups (SINR), it is defined as：

Wherein P represents the signal power that base station provides, and in the method, signal power is evenly distributed to all users. Represent user g_sPrecoding vector, be pre-coding matrix P_gS row,WithDone between the group for representing user respectively Interference, is respectively defined as in disturbing and organizing

Then user g_sSpeedCan be calculated by following formula respectively with speed C with system

As shown in figure 3, experimental result shows, compared with K- averages and K- CENTER ALGORITHMs, this method need not be specified in advance Packet count, can find out the cluster of arbitrary shape and size, can obtain more preferably group result with cancelling noise point, from And obtain the increase of preferable system conjunction rate.

Due to having tens to hundreds of antenna in transmitting terminal and receiving terminal, extensive mimo system can obtain very high Spectrum efficiency and energy efficiency, also therefore, extensive mimo system is increasingly paid close attention to as 5G technologies by people.But For Frequency Division Multiplex (FDD) system, CSIT feedback quantities can be increased dramatically with the increase of antenna number in extensive mimo system, This has become extensive mimo system bottleneck in actual applications.In order to tackle in extensive mimo system due to descending Estimate and uplink feedback takes the problem that resource excessively triggers, Ansuman Adhikary et al. and proposes a joint space Divide set multiplexing (JSDM) scheme.The main thought of the program has 2：User grouping and two grades of precodings.First, it is to carry The eigenmatrix of each subscriber channel covariance is taken, and all users point of similar features matrix will be possessed at same group, be then System uses two grades of precodings, packet transaction user.The JSDM schemes greatly reduce system dimensions, therefore reduce descending training With the consumption of uplink feedback.On the basis of this scheme, the present invention proposes the user grouping scheme based on density, the present invention The metric form of similarity between the eigenmatrix for measuring each subscriber channel covariance is first proposed, is then calculated using DBSCAN Method, using the Density Distribution between user by all user groupings.The advantage of DBSCAN algorithms is that can be divided into user appointing The group of what shapes and sizes, and it is not only circle, while the Density Distribution of user is estimated present invention employs k-dist figures, And determine therefrom that 2 important parameters of DBSCAN algorithms：Radius and minimum number of users threshold value.This programme improves system user point The method of group, has obtained more preferable throughput of system.

Claims (4)

1. density user packet method is based in a kind of extensive MIMO downlinks, it is characterised in that comprised the following steps：

Step 1, all users distance each other is calculated, obtain distance matrix；

Step 2, obtained according to distance matrix k-dist figure, according to the user density distributed intelligence included in figure obtain DBSCAN calculate Two important parameters of method：Radius and minimum number of users threshold value；

Step 3, cluster just is proceeded by from any one user later according to described two parameters, obtain final user grouping As a result.

2. density user packet method is based in extensive MIMO downlinks as claimed in claim 1, it is characterised in that step Rapid 1 is specially：The distance between any two user is calculated, the distance matrix dis of K × K dimensions is obtained, (k, j) is individual in matrix The distance between element representation user k and user j, are expressed as [dis]_k,j=S (U_k,U_j), wherein, S (U_k,U_j) according to be based on Distance between the user that the Similarity Measures of subspace projection are calculated.

3. density user packet method is based in extensive MIMO downlinks as claimed in claim 1, it is characterised in that step Rapid 2 are specially：If it the distance between is user j with k-th his nearest user that the k-dist of user j is represented, obtaining every After the k-dist values of individual user, by the arrangement of these value ascending orders, k-dist figures are then drawn as, the k-dist figures for obtaining contain user Density Distribution information, first point in the k-dist figures in first trough is threshold point；Wherein, in k-dist figures K values are equal to Minpts, and the value of the corresponding y-axis of the threshold point is exactly the eps values for obtaining.

4. density user packet method is based in extensive MIMO downlinks as claimed in claim 1, it is characterised in that step Rapid 3 are specially：First randomly choose a user k as initial user, and travel through it is all it is unmarked be the use of " processed " Family, finding out can obtain the eps- neighborhoods N of user k from the reachable user's set of the direct density of user k_eps(k), if | N_eps (k) |=1, then user k is exactly noise user, " processed " is marked the user as, then never it is labeled as " processed " again User in randomly choose next user；If 1 ＜ | N_eps(k) | ＜ Minpts, then user k is exactly edge customer, not right User k carries out any operation, but is never labeled as randomly choosing next user in the user of " processed "；If | N_eps (k) | >=Minpts, then user k is exactly core customer, it is possible to start a cluster for group；Gather since core customer k One group of class, N_epsK () is exactly a part for the group, then again from N_epsCore customer is looked in all users in (k), if deposited In core customer q, then by the eps- neighborhoods N of the core customer_epsQ all users in () are also grouped into the group, by these users It is collectively labeled as " processed ", and core customer is searched for from newly user is added to, repeat the above steps, constantly cluster, until Untill not having new core customer, all users for adding the group just constitute a user's group；After one group cluster terminates, then Never labeled as a user is randomly choosed in the user of " processed ", continue above-mentioned all steps, until all users all " processed " is marked as, at this time all users or noise user, or is divided into group, so far, user The packet stage just finishes；

After user grouping terminates, each group is worth to each center association side for organizing by seeking the covariance of all users in group Difference, i.e.,Then obtained between switched-beam matrix B, elimination group after interference using approximate BD algorithms, utilized MAX user's selection algorithm, user's selection is carried out to each group respectively, finally obtains pre-coding matrix P with ZF precoding algorithms；

After user's selection terminates, it is assumed that have S in g groups_gIndividual user is selected can to transmit data, thenRepresent whole The quantity of all users being selected in individual cell, usesS-th instant Signal to Interference plus Noise Ratio (SINR) of user in g groups is represented, It is defined as：

${\mathrm{\γ}}_{g_s}=\frac{\frac{P}{S}|h_{g_s}^H{B}_g{p}_{g_s}{|}^2}{1+{\mathrm{IAGI}}_{g_s}+{\mathrm{INGI}}_{g_s}}$

Wherein P represents the signal power that base station provides,Represent user g_sPrecoding vector, be pre-coding matrix P_gS Row,WithInterference and the interior interference of group between the group of user are represented respectively, are respectively defined as

${\mathrm{IAGI}}_{g_s}=\underset{t\≠s}{\Σ}\frac{P}{S}|h_{g_s}^H{B}_g{p}_{g_t}{|}^2$

${\mathrm{INGI}}_{g_s}=\underset{g^{\′}\≠g}{\Σ}\underset{t=1}{\overset{S_{g^{\′}}}{\Σ}}\frac{P}{S}|h_{g_s}^H{B}_{g^{\′}}{p}_{{g_t}^{\′}}{|}^2$

Then user g_sSpeedCan be calculated by following formula respectively with speed C with system

${\mathrm{\χ}}_{g_s}={\log }_2(1+{\mathrm{\γ}}_{g_s})$

$C=\underset{g=1}{\overset{G}{\Σ}}\underset{s=1}{\overset{S_g}{\Σ}}{\mathrm{\χ}}_{g_s}.$