METHOD AND APPARATUS FOR CHANNEL PREDICTION FOR 5G UPLINK/DOWNLINK MASSIVE MIMO SYSTEM FOR OPEN RADIO ACCESS NETWORKS

Abstract

A method for channel prediction for uplink (UL) and downlink (DL) massive Multiple Input Multiple Output (MIMO) systems for Open Radio Access Network (0-RAN) fronthaul Split 7.2 networks enables prediction of a channel that is seen by the UL slot. The pre-processing matrix is computed by the distributed unit (DU) based on this predicted channel and sent to the radio unit (RU) for minimizing the effects of channel aging. A channel corresponding to sounding reference signal (SRS) symbol closest to uplink slot being decoded can be predicted from previous SRS symbols and can be used as a combining matrix. Alternatively, the channel of the uplink slot itself can be predicted from past SRS symbols, and a combining matrix can be generated based on the predicted channel.

Description

BACKGROUND OF THE DISCLOSURE

1. Field of the Disclosure

The present disclosure relates to a method and apparatus for Radio Access Networks (RANs), and relates more particularly to a method and apparatus for mitigating channel aging in Open RAN (O-RAN).

2. Description of the Related Art

Conventional RANs were built employing an integrated unit where the entire RAN was processed. Conventional RANs implement the protocol stack (e.g., Physical Layer (PHY), Media Access Control (MAC), Radio Link Control (RLC), Packet Data Convergence Control (PDCP) layers) at the base station (also referred to as the evolved node B (eNodeB or eNB) for 4G LTE or next generation node B (gNodeB or gNB) for 5G NR). In addition, conventional RANs use application specific hardware for processing, which make the conventional RANs difficult to upgrade and evolve. As future networks evolve to have massive densification of networks to support increased capacity requirements, there is a growing need to reduce the capital costs (CAPEX) and operating costs (OPEX) of RAN deployment and make the solution scalable and easy to upgrade.

Cloud-based Radio Access Networks (C-RANs) are networks where a significant portion of the RAN layer processing is performed at a baseband unit (BBU), located in the cloud on commercial off the shelf servers, while the radio frequency (RF) and real-time critical functions can be processed in the remote radio unit (RRU), also referred to as the radio unit (RU). The BBU can be split into two parts: centralized unit (CU) and distributed unit (DU). CUs are usually located in the cloud on commercial off the shelf servers, while DUs can be distributed. The BBU may also be virtualized, in which case it is also known as vBBU. Radio Frequency (RF) interface and real-time critical functions can be processed in the Radio Unit (RU), sometimes referred to as Remote Radio Unit (RRU).

The O-RAN architecture is a Cloud-based architecture specified by the 0-RAN Alliance. The components of the O-RAN architecture include, e.g., the Service Management and Orchestrator (SMO) Framework, the Non-Real Time (Near-RT) Radio Intelligent Controller (RIC), the Near-Real Time (Near-RT) Radio Intelligent Controller (RIC), the O-RAN Centralized Unit (0-CU), the O-RAN Distributed Unit (0-DU), and the O-RAN Radio Unit (0-RU). The data between O-DU and O-RU is sent over the open fronthaul interface.

As noted above, C-RAN connects many RUs, each catering to a 4G/5G cell, to a central baseband processing unit via a fronthaul. This provides performance benefits via inter-cell and inter-frequency co-ordination and also cost benefits through resource pooling and reduced installation space. However, the common public radio interface (CPRI) used in C-RAN does not sufficiently prescribe specifications for fronthaul interfaces, thereby resulting in different specifications for different vendors. The drawback is lack of inter-operability between equipment from different vendors. The Open RAN (O-RAN) fronthaul specifications were formulated against this background and aim to help make multi-vendor RAN a reality. Furthermore, due to bandwidth limitations of the fronthaul, 0-RAN includes different fronthaul functional splitting of the physical layer, e.g., fronthaul Split 7-2 and Split 7-3.

As shown in FIG. 1, in Split 7-2 the RU 101 (handling lower physical layer (LPHY)) receives analog signals from radio frequency (RF) units, converts the signals to digital signals (shown by the function block 1001 designated “RF/ADC”, indicating conversion of analog signals from the RF units to digital signals), does time domain processing, applies FFT (as well as performing cyclic prefix (CP) removal) and sends the frequency domain signal to the DU, over the fronthaul 103. Additional functions shown on the RU side include: SRS over all antennas; combining matrix buffer; and combining matrix application on UL. The DU 102 processes signals of the RUs, e.g., RU 101. Functions shown on the DU side include: SRS processing for channel estimation, user pairing and weight computation; rest of L1 processing; and L2 processing. In uplink massive MIMO, each RU is equipped with a large number (represented by variable NR) of antennas, e.g., typically upwards of 64. In uplink multi-user-MIMO (MU-MIMO), NU number of user equipments (UEs) transmit on the same time-frequency resource. The fronthaul 103 between the RU 101 and DU 102 has limited bandwidth and what is sent over it has to be designed judiciously. The RU combines the NR values received over NR antennas per subcarrier (after FFT) into M NR values by applying a M xN_R combining matrix and sends only M values per subcarrier over the fronthaul to the DU. The DU decodes the NU UEs from these Mvalues. The combining matrix for an uplink slot needs to be informed by DU apriori before the start of the reception of an uplink slot at RU. This processing is for physical uplink shared channel (PUSCH) channel. However, for sounding reference signal (SRS) channel, all N_R values are sent to DU for every subcarrier. In Split 7-3, there is no fronthaul and all processing (RF/ADC conversion; FFT and CP removal; UL MU-MIMO docoder processing; rest of Li processing; and L2 processing) is in one place, as shown in the upper portion of FIG. 1.

For mobility use cases (typically 30 km/hr), using conventional methods for the Split 7.2 architecture results in a loss of performance in comparison to Split 7.3 method. This is because of the channel aging issue associated with the combining matrix, which is discussed in further detail below. Accordingly, there is a need to mitigate the channel aging issue for the Split 7.2.

SUMMARY OF THE DISCLOSURE

An example embodiment of the apparatus and method according to the present disclosure provides a mechanism to predict the channel at desired time instant in massive MIMO systems (e.g., utilizing O-RAN architecture) so that the precoding matrix in the downlink and combining matrix in the uplink are derived from it to mitigate the channel aging.

An example embodiment of the method according to the present disclosure provides a training period in which the prediction coefficients are obtained by training with the latest Sounding Reference Signal (SRS) symbol as the desired channel and the past SRS symbols as the inputs to the prediction module. Following the training period, the prediction coefficients are used to predict the future SRS symbols by using the past SRS symbols. The predicted SRS symbol is used to derive the combining matrix to be applied in RU for decoding the UL slot close to the slot containing SRS symbols.

An example embodiment of the method according to the present disclosure includes: i) predicting the channel between the periodic SRS symbols, and using the predicted channel to decode the UL slots anywhere between the slots containing SRS symbols, and ii) reconstructing the predicted channel between two SRS channels/symbols.

In an example embodiment of the method according to the present disclosure, the channel of intra-site interferers is predicted and used to perform interference whitening for decoding the desired users.

An example method according to the present disclosure is applicable for precoding in the DL as well using the predicted channels.

According to an example embodiment, the method according to the present disclosure can be implemented at least in part using adaptive filters.

According to an example embodiment, the method according to the present disclosure can be implemented at least in part using Kalman filters.

According to an example method according to the present disclosure, SRS symbols are used to determine speed and replace the learning phase of linear channel prediction.

An example method according to the present disclosure utilizes Delay-Doppler domain to perform channel prediction of time domain multipath taps, thereby enabling prediction of channels in a future SRS/PUSCH symbol in any RB, which method is conducive for sub-band SRS scheduling and increases the SRS capacity.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 illustrates the architecture of O-RAN Split 7.3 and Split 7.2.

FIG. 2 illustrates various processing and scheduling associated with reception of an uplink slot in a DSUUU slot pattern.

FIG. 3 illustrates three physical sites and their cells/sectors classified as desired, intra-site and inter-site cells/sectors.

FIG. 4 illustrates the effect of using predictions of both LCP methods vs. not using the LCP methods.

FIG. 5 illustrates channel reconstruction error associated with Nyquist sampling/reconstruction of LCP Method 2.

FIG. 6 illustrates Split 7.3 and Split 7.2 having same performance and compares it with other methods for FUS.

FIG. 7 illustrates that LCP Method 2 overcomes limitations of LCP Method 1 for TUS.

FIG. 8 illustrates the comparisons between LCP Methods 1 and 2 for FUS and TUS.

DETAILED DESCRIPTION

The present disclosure provides an improved mechanism to overcome the loss of performance arising from the channel aging issue. An example embodiment of a method according to the present disclosure, based on linear channel prediction, ensures that Split 7.2 and 7.3 achieve substantially the same performance.

The present disclosure provides a mechanism for channel prediction for uplink and downlink massive MIMO systems for O-RAN fronthaul Split 7.2 networks. In traditional networks, which include co-located RU and DU, although the advantages of ORAN fronthaul split 7.2 are not available, the channel estimates can be based on the DMRS present in the UL slots, and hence the channel aging problem is not an issue. On the other hand, in O-RAN fronthaul Split 7.2 networks, because of the separation of RU and DU, UL slots have to be pre-processed in the RU before sending to DU to minimize the FH requirements. This means the last estimated SRS channel is used in the DU to determine the combining weights which are sent to RU for pre-processing the incoming UL slots, thereby resulting in performance degradation due to channel aging.

In the present disclosure, channel prediction techniques are provided to predict the channel that is seen by the UL slot. The pre-processing matrix is computed by the DU based on this predicted channel and sent to RU for minimizing the effects of channel aging. In an example embodiment, a channel corresponding to SRS symbol closest to uplink slot being decoded is predicted from previous SRS symbols and is used as a combining matrix. In yet another example embodiment, the channel of the uplink slot itself is predicted from past SRS symbols, and a combining matrix is built based on the predicted channel. Low-complexity, implementation-friendly versions of the linear predictor is implemented using adaptive filters, Kalman filters and/or estimating speed of users via SRS symbols.

As an example, a pattern of repeated slots is considered, denoted by DSUUU, where D, S and U are downlink, special, and uplink slots, respectively. A downlink slot has only downlink OFDM symbols, and an uplink slot has only uplink OFDM symbols. A special slot has a combination of downlink, flexible and uplink OFDM symbols, in that order. The quantities D,, S,, U, denote downlink, special and uplink slots, respectively, at a slot index t. The uplink OFDM symbols towards the end of a special slot carry SRS of various multiplexed users across the entire bandwidth. As an example, 30 kHz subcarrier spacing is used, which means a slot is 0.5 ms in duration and SRS periodicity is five slots or 2.5 ms. Referring to the DSUUU slot pattern shown in FIG. 2 (which is repeated three times in FIG. 2), in order to prepare for the reception of uplink Slot U_t+1 after the special slot S_t-5, the DU computes the channel estimates of all users across the entire bandwidth. The DU then decides which N_U users need to be paired for transmission in the same time-frequency resource, and then computes the combining matrix to receive these N_U users over a selected RB in Slot U,+i. This combining matrix is sent to the RU before the reception of U,+i. As shown in FIG. 2, the sequence of first seven slots, collectively referenced as 201, is used for full band channel sounding in special slots. Also shown in FIG. 2 is a second sequence of UUU slots, collectively referenced as 202, which slots are used for computing channel estimates, user paring and combining matrix based on scheduled users who are scheduled for reception in U,+i and transfer combining matrix to the RU before U_t+1.

In FIG. 3, three physical sites are shown, ABCDEF, CKLMND and IJAFGH. Each physical site is comprised of three sectors, e.g., as shown by the physical site ABCDEF in FIG. 3, which has three sectors OABC, OCDE and OEFA. Each sector is considered to be a cell, so “sector” and “cell” will be used interchangeably. The base station (RUs and DU) for the physical site ABCDEF resides at the location indicated as O in FIG. 3. The DU receives signals from three RUs, each corresponding to a sector/cell within the physical site ABCDEF. Signals of users in sector/cell OABC (referred to as desired users) received at the base station at O have interference from users in other sectors/cells (OCDE and OEFA) of the same physical site, which interference is called intra-site interference. The desired user also experiences interference from users in sectors/cells of other physical cell sites (CKLMND and IJAFGH), which interference is called inter-site interference. Because the DU of a physical site has access to scheduling information of users in all the sectors/cites belonging to the same physical site, the DU at O can estimate the channels of the desired users and intra-site interferers, but the DU can't estimate the channels of inter-site interferers. Intra-site and inter-site interferers are collectively abbreviated as ICI (Inter-cell interference). The desired user at the center of a cell will have negligible ICI, which can be neglected, while the desired user at the cell edge will experience significant ICI. In the present disclosure, the case of absence of ICI (e.g., for cell-center desired user) is first analyzed, and then the case of presence of ICI (e.g., for cell-edge desired user) is analyzed.

Brief explanation of notations is provided here. The quantity E {x}denotes the expectation of x. The quantity I represents an Identity matrix of appropriate dimension. The estimate of x is denoted by {circumflex over (x)}. Matrices and vectors will be represented by bold uppercase and lowercase, respectively, while scalars are denoted by normal fonts. Matlab notation is used to access parts of matrices/vectors. The (a,b)^th element of a matrix X is denoted by X(a,b). The p element of a vector x is denoted by x(p). The quantity [xJ denotes a floor operation on x, i.e., it denotes the greatest integer less than x. As an example, [4.19]=4. The quantity x*is the conjugate of x.

SYSTEM MODEL IN THE ABSENCE OF ICI

For this scenario, we are concerned with demodulating OFDM data symbols in uplink Slot U_tm+i which is i slots away from nearest special slot St that contains the SRS. Let the received signal in a given OFDM data symbol in uplink Slot U_tm+iacross the N_R antennas on a given subcarrier s in a given Resource block (RB) rbe denoted by the N_R ×1 vector y (for the sake of simplicity, we do not parametrize y, H, x and n by the OFDM data symbol index and Parameters s, r, and t_m+i),

y=Hx+n  (1)

where H is a N_R ×N_U channel matrix of desired users, x is N_U x 1 vector of desired users, n is additive white Gaussian noise (AWGN) whose covariance matrix is C=E{nnH}=ag. The N_R ×N_U channel matrix, H_DMRS, is the average of all H across subcarriers corresponding to the given RB rand the first demodulation reference symbol (DMRS) in an uplink Slot U_tm+i. Let the N_R ×N_U channel matrix corresponding to a given Subcarrier s in the given RB r in an SRS symbol in special Slot S_tm−5(SRS periodicity is five slots) be denoted by Hi. The quantity H_SRS denotes the average of all Hiacross subcarriers in the given RB r. The estimate of channel matrix H for all subcarriers in the given RB rand all OFDM data symbols in uplink Slot U_tm+iis either HDMRS² or H_SRS.

Two points should be noted regarding the above-described system model in the absence of ICI. First, in 5G N_R, an RB is a set of 12 subcarriers. In the present disclosure, an RB is a set of 12 subcarriers in any OFDM symbol (as per Section 4.4.4.1 in 3GPP TS 38.211, “Physical Channels and Modulation,” 3GPP, V15.8.0, Dec 2019). The RBs in any OFDM symbol are indexed from r=0, . . . , 49, as we use 50 RBs across the channel bandwidth in the present disclosure. Second, channel estimates for OFDM data symbols in any subcarrier can be interpolated from channel estimates of all DMRS in the slot for that subcarrier. However, for the sake of simplicity, channel estimates of all OFDM symbols in the slot can be based on the channel estimate derived in the first DMRS only.

The O-RAN fronthaul Split 7.2 and Split 7.3 receivers are defined as follows:

• Split 7.3: An estimate of {circumflex over (x)} is computed as {circumflex over (x)}=(H^H_DMRSH_DMRS+σ²I)⁻¹ H^H_DMRSY
• Split 7.2: An M xN_R combining matrix W_RU is used by RU to compress y.
• What we get at the DU is

y=Hx+n  (2)

where y=W_RUy, H=W_RUH and n=W_RUn. We use one combining matrix for all subcarriers in an RB across the OFDM data symbols in Uplink slot U_tm+i. At the DU, the effective channel estimate used for all subcarriers in the given RB r across OFDM data symbols in uplink Slot U_tm+iis 1=W_RUH_DMRS and the covariance matrix seen by DU is C =E{nn^H}=σ²W_RUW_RU^H . The DU receives y and computes an estimate of x as {circumflex over (x)} =W_DUy where the minimum mean square error (MMSE) solution for W_DU is given as:

W _DU=(HDMRSHC−1HDMRS+I)−1HDMRSHC−1.  (3a)

or the alternate form

W _DU =H _DMRS ^H(H_DMRSH_DMRS^H+C)⁻¹  (3b)

Note that if WRu=HiiktRs (conjugate combining matrix or CCM), we have

W_DU==(H^H_DMRS H_DMRS+σ²I)⁻¹  (4)

which translates to estimating x as

{circumflex over (x)}=W_DUW_RUy==(H^H_DMRS H_DMRS+σ²I)⁻¹ H^H_DMRSy  (5)

which is the MMSE of estimate x based on y. One can select W_RU=(H^HDMRS H_DMRS+σ² I)⁻¹ H^H_DMRS, called as MMSE combining matrix or MMSE-CM. In this case, no processing is required in DU and W_DU=I.

In this section, signal-to-interference-plus-noise ratio (SINR) computation will be discussed. If an estimate of {circumflex over (x)} is computed as {circumflex over (x)}=Wy={tilde over (H)}x+ñ, where {tilde over (H)} =WH and if ñ=Wn, the SINR of the ith user is given as

$\begin{array}{cc}{\mathrm{SINR}}_i=\frac{{❘\tilde{H}\left(i,i\right)❘}^2}{\sum _{j=1,j\ne i}^{N_U}{❘\tilde{H}\left(i,j\right)❘}^2+{\sigma }^2\sum _{j=1}^{N_U}{❘W\left(i,j\right)❘}^2}& \left(6\right)\end{array}$

Furthermore, we define the spectral mean of a set of SNRs as follows. Suppose there are N_USINRs, i.e., SINR₁, . . . ,SINR_Nu, the spectral mean (SM) SINR of these SINRs is given by

$\begin{array}{cc}{\mathrm{SINR}}_{\mathrm{SM}}=\sqrt[N_U]{\left(1+{\mathrm{SINR}}_1\right)\dots \left(1+{\mathrm{SINR}}_{N_U}\right)}-1·& \left(7\right)\end{array}$

The SINRsM of a set of N_U SNRs is the SINR which will give the same spectral efficiency (as the sum of spectral efficiencies of the corresponding N_U SINRs) if each of the Nu SINRs is replaced by SINRsM.

The above-noted MMSE equation (5) is implementable in the Split 7.3 architecture as the uplink Slot U,is decoded after being received over air and the Split 7.3 architecture can compute H_DMRS. While the CCM and MMSE-CM methods also reduce to the MMSE solution, CCM and MMSE-CM can't be implemented in Split 7.2. This is because the combining matrix W_RU that needs to be sent to RU before the uplink Slot U.+, can't be based on H_DMRS as DU is yet to have access to it (one needs uplink Slot U.+, to compute it). As an alternative, approximations of CCM and MMSE-CM methods can be used in the context of Split 7.2. The approximations of the two methods are obtained by replacing H^HDMRS with H^HSRSin WR_U and are given by the following: CCM approximation to Split 7.2: The combining matrix is W_RU=H!n The quantity W_DU is as per above-noted equation (3).

MMS1E-CM approximation to Split 7.2: The combining matrix is W_RU=(H^HS_RS HS_RS+c²I ) H^HsRs. The quantity W_DU is the Identity matrix I.

Note that the approximations of CCM and MMSE-CM to Split 7.2 will incur a loss of performance relative to Split 7.3 (which uses MMSE method) at high speeds. At high speeds, there is a significant difference between HsRS and H_DMRS as they are channel estimates of channels that are 5+i slots apart (Sm-s and Um+,, respectively). This is called channel aging effect and depends on Doppler spreadfo. The Doppler spread depends on speed of the UE v and carrier frequency fcasfv=fc(v/c), in which c is the speed of light. Accordingly, at high speeds Split 7.2 CCM and MMSE methods will have a loss of performance compared to Split 7.3. Split 7.3 can be treated as the upper bound for Split 7.2. At low speeds, Split 7.2 tends to be the same as Split 7.3 or the MMSE method as the channel aging is very minimal. One of the goals of the method according to the present disclosure is how to design the combining matrices such that the gap between a Split 7.2 and 7.3 is reduced.

An example embodiment of the method according to the present disclosure uses linear channel prediction (LCP) to address this loss of performance of Split 7.2 relative to Split 7.3 due to channel aging. The combining matrix has to be sent to RU before the reception of uplink Slot U_tm+i ,+,. The DU can use SRS channel estimates of Slots Sm-s and the ones before that to build the combining matrix. Note that the DU can't use SRS slot Sm, due to latency requirements and the SRS slot being close to uplink slot Um+,. The LCP method in the DU uses channel estimates of SRS in Slots S_M,-s and ones before that to predict an estimate of H_DMRS (that corresponds to Um.,) as H_DMRS and uses this in CCM-LCP (Split 7.2) and MMSECM-LCP (Split 7.2) methods as W_RU=H_DMRS and W_RU=(H_DMRSHD_MRS+Q^ZI)-iH_DMRS, respectively. The quantity WD_U is as per above-noted equation (3) for CCM-LCP (Split 7.2) and the Identity matrix I for MMSE-CM-LCP (Split 7.2).

LCP methods for Split 7.2 in the absence of ICI will be discussed below. LCP is generally used to predict future values of a channel based on present and past values. The LCP equation can be written as S_up_u=d_u(will be defined later). LCP operates in two modes. First, in learning/training mode, we populate the input matrix S. and desired vector d_u. Once populated we determine the LCP coefficient vector p_u. Next is the prediction mode, where we populate the input Matrix S. based on present and past values and use the already-learnt LCP coefficient vector _pu to compute/predict vector of future channel values d_u. There are two methods of LCP for Split 7.2 that will be discussed below.

LCP Method 1: This method uses channel estimates of only SRS symbols to learn the LCP coefficients. Define θ_i.t.r.a as the channel estimate of User i in RB r, antenna a and channel corresponding to SRS symbol in a special Slot S_t. Likewise, ω_i.t.r.a as the channel estimate of User i in RB r, antenna a and channel corresponding to the first DMRS symbol in an uplink Slot U_t. The quantities θ_i.t.r.a and ω_i.t.r.a can be considered as the average of the channel values of all subcarriers in a RB. We now discuss the learning/training mode for User u at Special slot S,. We now build an input matrix S_r.t.u as

$\begin{array}{cc}{S}_{r,t,u,}=\left[\begin{array}{cccc}{\theta }_{u,t-5-5\left(o_t-1\right),r,1}& \dots & {\theta }_{u,t-10,r,1}& {\theta }_{u,t-5,r,1}\\ {\theta }_{u,t-5-5\left(o_t-1\right),}r,2& & {\theta }_{u,t-10,r,2}& {\theta }_{u,t-5,r,2}\\ ⋮& ⋮& ⋮& ⋮\\ {\theta }_{u,t-5-5\left(o_t-1\right),r,N_R}& \dots & {\theta }_{u,t-10,r,N_R}& {\theta }_{u,t-5,r,N_r}\end{array}\right]& \left(8\right)\end{array}$

and desired/predicted vector d_r,t,u as

$\begin{array}{cc}{d}_{r,t,u}=\left[\begin{array}{c}{\theta }_{u,t,r,1}\\ {\theta }_{u,t,r,2}\\ ⋮\\ {\theta }_{u,t,r,N_R}\end{array}\right].& \left(9\right)\end{array}$

The preliminary LCP equation is S_r,t,up_u=d_r.t.u. This corresponds to using past values in RB rto predict future values in same RB r. In the preliminary LCP equation, one can use values in RBs r - o_f, . . . ,r, . . . ,r+of to predict future values in RB r where of is the frequency-domain prediction order (either side of RB_r). The equation would then be [srr−oj,i,a. . . Sr_,t,u^T . . . S_r+o,j,t,a^T]^Tp_a=d_r,t,a. Frequency-domain prediction order needs to be employed if there is correlation across RBs in frequency domain and this will generally exist for low delay spreads which result in frequency-selectivity across larger number of RBs in frequency domain. However, for the sake of simplicity, we consider of =0 in the present disclosure. Note that each row of the input matrix S_r,t,u has o_t channel estimates and we say that order of prediction is o_t. Each row of S_r,t,u corresponds to an unique antenna and RB r. One can stack S_r,t,u and d_r,t,u for various time instants t=t₁, . . . ,t_n as

$\begin{array}{cc}{S}_{r,u}=\left[\begin{array}{c}{S}_{r,t_1,u}\\ ⋮\\ S_{r,t_u,u}\end{array}\right],d_{r,u}=\left[\begin{array}{c}{d}_{r,t_1,u}\\ ⋮\\ d_{r,t_u,u}\end{array}\right]& \left(10\right)\end{array}$

Furthermore, we can stack the various S_r,u and d_r,u for RBs r₁, . . . ,r_A as

$\begin{array}{cc}{S}_u=\left[\begin{array}{c}{S}_{r_1,u}\\ ⋮\\ S_{r_A,u}\end{array}\right],d_u=\left[\begin{array}{c}{d}_{r_1,u}\\ ⋮\\ d_{r_A,u}\end{array}\right].& \left(11\right)\end{array}$

We now have the relation

S_up_u=d_u  (12)

which is the LCP equation used for training User u (determining LCP coefficient vector p_u using SRS in slots across time and RBs). The estimate of p_u is obtained as p_u =(S_u^HS_u)⁻¹ S_u^Hd_u.

Once the estimate of LCP coefficient vector pu is determined, we can use it in the prediction mode. Let it be required to decode an uplink Slot U_tm+i in RB r where i=1,2,3. As shown in FIG. 2, we are dealing with the slot pattern DSUUU with SRS periodicity of five slots. The nearest special slot to U_tm+i containing an SRS symbol is S_tm. Further let us assume t_m>t_n where t_n is the last special slot containing SRS used in the learning phase. We compute an estimate of d_r,tm,u as

{circumflex over (d)}_r,tm,u=S_r,tm,u{circumflex over (P)}_u  (13)

The elements of {circumflex over (d)}_r,tm,u will denote channel estimates of User u in SRS of Slot S_tm for various antennas. We use this channel estimate as the channel estimate of the DMRS in any of the next three uplink slots U_tm+i ,+,,i=1,2,3. While this would be ok for the uplink slot U_tm+1 immediately after the special slot Sm, (referred to as first uplink slot after special slot or FUS), there would be considerable loss of performance for the third uplink slot U_tm+3 after the special slot S_tm (TUS). The LCP Method 2 discussed below addresses this issue of LCP Method 1. An estimate of H_DMRS that corresponds to uplink slots U_tm+i,i =1,2,3 and RB r is given as

Ĥ_DMRS(p, q)={circumflex over (ω)}_q,tm+i,r,p ≈{circumflex over (d)}_r,tm,q(p)  (14)

which can then be used in combining matrix of CCM-LCP (Split 7.2) or MMSE-CM-LCP (Split 7.2). Note that though H_DMRS corresponds to uplink Slot U_tm+i ,+,,i=1,2,3, the calculation can be done any time after special Slot Sm-s and can be sent to RU well ahead of reception of uplink Slot U_tm+i,i=1,2,3. If H_DMRS ≈H_DMRS then we can expect Split 7.2 to perform as well as Split 7.3 (which is what we will see later in this disclosure).

LCP Method 2: As described above, LCP Method 1 only predicts the channel of a user in SRS symbol of a special slot closest to the uplink slot. While this is ok for FUS, there is loss of performance for TUS. LCP Method 2 addresses this issue by reconstructing the channel between two SRS symbols and using the reconstructed channel to build the desired vector d,r..u. In order to perfectly reconstruct the channel between two SRS symbols/special slots, the channel should be sampled at a frequency greater than the Nyquist sampling frequency, which is²fD. Consequently, the distance between two special slots should be less than 2AD which is the Nyquist sampling period, which means reconstruction is possible up to a speed of v=61.7 km/hr atfc=3.5 GHz carrier frequency.

Let it be required to reconstruct the channel (or the channel estimate) ω_u,t+i,r,a uplink Slot U_t+i,i=1,2,3 where the nearest special slot preceding it is S,. As per Whittaker-Shanon interpolation and Nyquist-Shannon sampling theorem

$\begin{array}{cc}{\hat{\omega }}_{u,t+i,r,a}=\sum _{n=-\infty }^{\infty }{\theta }_{u,t+5n,r,a}\mathrm{sinc}\left(\frac{d_i-{\mathrm{nT}}_{\mathrm{SRS}}}{T_{\mathrm{SRS}}}\right)& \left(15\right)\end{array}$

where d₁ is the time between the starts of the SRS symbol in Special slot S_t and first DMRS in uplink Slot U_t+i, T_SRS=25 ms is the SRS periodicity. Note that to compute the estimate of {circumflex over (ω)}_u,t+i,r,a, infinite past and future values of θ_u,t+5n,r,a are needed, but we approximate the above summation by using only B₁ past and B₂ future values. It should be noted that the Nyquist-Shannon sampling theorem is for reconstruction in the time domain, but since time-domain to frequency-domain is a linear transformation, the reconstruction can be applied in frequency-domain as well.

We now discuss the learning/training mode, which can be done only at or after Special slot S_t+5B2. The desired/predicted vector d_r,t,+i,u is constructed as

$\begin{array}{cc}{\overline{d}}_{r,t+i,u}=\left[\begin{array}{c}{\hat{\omega }}_{u,t+i,r,1}\\ {\hat{\omega }}_{u,t+i,r,2}\\ ⋮\\ {\hat{\omega }}_{u,t+i,r,N_R}\end{array}\right].& \left(16\right)\end{array}$

The preliminary LCP equation is Sr,tu,pu=d_r,t+i,u. We retain the same definitions for S_r,t,u,S_r,u,S_u. as discussed above in connection with LCP Method 1, and we i) stack the various d_r,t+i,u for time instants t=t₁,t₂, . . . ,t_n, to arrive at d_r,u and ii) then stack the various d_r,u for RBs r₁, . . . ,r_A to arrive at d_u as

$\begin{array}{cc}\begin{array}{cc}{\stackrel{\_}{d}}_{r,u}=\left[\begin{array}{c}{\overline{d}}_{r,t_1+i,u}\\ ⋮\\ {\overline{d}}_{r,t_n+i,u}\end{array}\right]& {\stackrel{\_}{d}}_u=\left[\begin{array}{c}{\overline{d}}_{r_1,u}\\ ⋮\\ {\overline{d}}_{r_A,u}\end{array}\right]\end{array}& \left(17\right)\end{array}$

We now have the relation

S_up_u=d_u  (18)

which is the LCP equation used for training User u (determining LCP coefficient vector p_u).

The estimate of p_u is obtained as {circumflex over (p)}_u=(S_u^HS_u )⁻¹ S_u^H{circumflex over (d)}_u . Once LCP coefficient vector p_u is determined, we can use it in the prediction mode. Let it be required to decode an uplink Slot U_tm+i, in RB r where i=1,2,3. As shown in FIG. 2, the slot pattern DSUUU with SRS periodicity of five slots is used. The nearest special slot to U,+, containing an SRS symbol is S_tm . Furthermore, it is assumed t_m>t_n+B₂ where t_n is the last special slot containing SRS used in the learning phase. This will allow for the channel reconstruction at uplink Slot U_tn+i and also learning p_u using S_u, {circumflex over (d)}_u at time instants t₁, . . . ,t_n. We compute an estimate of {circumflex over (d)}_r,tm+iu as

{circumflex over (d)}_r,tm+i,u=S_{r,tm,u{circumflex over (P)}u}  (19)

The elements of {circumflex over (d)}r,t_m+i,u, are the channel estimates of User u in DMRS of uplink Slot U_tm+i,i=1,2,3, for various antennas. An estimate of H_DMRS that corresponds to first DMRS in uplink slots U_tm+i ,+,,i=1,2,3 and RB r is given as

Ĥ_DMRS(p,q)=ŵ_q,,tm+i,r,p=d_r,tm+i,q(p)  (20)

which can then be used in combining matrix of CCM-LCP (Split 7.2) or MMSE-CM-LCP (Split 7.2). Note that though H_DMRS corresponds to uplink Slot U_tm+i,i=1,2,3, the calculation can be done any time after special Slot S,M-s and can be sent to RU well ahead of reception of uplink Slot U_tm+i,i=1,2,3.

In the following section, example system models and LCP methods for Split 7.2 in the presence of ICI are discussed (along with MMSE receiver for Split 7.3). The channel of the ith desired user in a Subcarrier s in Resource block (RB) r, Slot I and Antenna a is denoted by _hi,t,s,a. If Slot t is a special slot, h_i,t,s,a corresponds to the SRS OFDM symbol and if it is an uplink slot, it corresponds to the first DMRS of the slot. Similar to the above-noted equation (1), let the received signal in a given OFDM data symbol in uplink Slot U_tm+i, across the N_R antennas on a given subcarrier s in a given Resource block (RB) r be denoted by the N_R ×1 vector y (for the sake of simplicity, we do not parametrize y, H, x, G, xi, J, x₂ and n by the OFDM data symbol index and parameters s, r, t_m+i)

y=Hx+Gx₁+Jx₂+n  (21)

where H is a N_R ×N_U channel matrix of Nudesired users, x is N_U x₁ vector of desired users, G is a N_R ×N_U1 channel matrix of Nui intra-site interfering users, x₁ is N_U1 ×1 vector of intra-site interfering users, J is a N_R ×N_U2 channel matrix of N_U2 inter-site interfering users, x₂ is N_U2 ×1 vector of inter-site interfering users, and Ĩ, defined in equation (22) below,

Ĩ=GX₁+JX₂+n  (22)

is the total interference. Likewise, let the N_R ×N_U1 channel matrix of intra-site interferers corresponding to a given Subcarrier s in the given RB r in a SRS symbol in special Slot S_tm-5 (SRS periodicity is five slots) be denoted by G₁. The DU receives signals from RUs of all three sectors/cells in a physical site, so it has access to intra-site channels G but not to inter-site channels J. The quantities G_DMRS and G_SRS correspond to channel matrix G and G₁ of intra-site interferers the same way as H_DMRS and H_SRS correspond to channel matrix H and H₁ as discussed above in connection with LCP methods for Split 7.2 in the absence of ICI.

In this section, MMSE receiver for Split 7.3 is discussed. The interference covariance matrix (ICV) C_7.3 is given as

$\begin{array}{cc}{C}_{7.3}=E\left\{\text{?}\right\}=\frac{1}{12}\sum _{\text{?}}\left(y-\mathrm{Hx}\right){\left(y-\mathrm{Hx}\right)}^H={\sigma }^{2}I+\frac{1}{12}\sum _{\text{?}}\left({\mathrm{GG}}^H+{\mathrm{JJ}}^H\right)& \left(23\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

where the expectation operator is implemented by averaging (y−Hx(y−Hx)^H over the 12 subcarriers in RB r of the first DMRS in uplink Slot U_tm+i. The quantity x denotes the data vector of desired users corresponding to any Subcarrier s in a RB r and any OFDM data symbol in uplink Slot U_tm+1. An estimate of x is given as

x=H_DMRSC₇₃⁻¹H_DMRS+I)⁻¹H_DMRS^HC_7.3⁻¹y  (24a)

or alternatively,

{circumflex over (x)} =H_DMRS^H(H_DMRSH_DMRS^H+C_7.3)⁻¹ y  (24b)

In this section, receivers for Split 7.2 are discussed. At the DU, for calculation of W_DU as per equation (3), we need to update C which is equal to W_RUC_7.3W_RU ^H and which is calculated by averaging (Y- Hx)(Y-Hx)^H over the 12 subcarriers in the RB corresponding to first DMRS in received uplink slot at DU. In the above section discussing the LCP methods for Split 7.2 in the absence of ICI, we defined θ_i,t,r,a as the channel estimate of desired User i in RB r, antenna a and channel corresponding to SRS symbol in a special Slot S_t. Likewise, we define δ_i,t,r,a as the channel estimate of intra-site interfering User i in RB r, antenna a and channel corresponding to SRS symbol in a special Slot S_t. The quantities θ_i,t,r,a and δ_i,t,r,a are considered as the average of the channel values of all subcarriers in a RB. Just as we computed Ĥ_DMRS corresponding to uplink Slot U_tm+i, in equations (14) and (20) (using LCP Methods 1 and 2 and θ_i,t,r,a), we similarly compute an estimate Ĝ_DMRScorresponding to uplink Slot U_tm+i , using δ_i,t,r,a . An estimate of ICM for Split 7.2 is then given by

Ĉ_7.2=Ĝ_DMRS^H+σ² I  (25)

which is just an approximation to ICV (ICVA) as we use average channel value per RB to compute ICV. A more accurate way of computing the ICV would be to use all channel values per subcarrier in the computation which is done as follows. The channel of the ith intra-site interfering user in a Subcarrier s in Resource block (RB) r, Slot i and Antenna a is denoted by g_i,t,s,a(note that the average of g_i,t,s,a for all subcarriers s in RB r is δ_i,t,r,a). If Slot t is a special slot, g_i,t,s,acorresponds to the SRS OFDM symbol and if it is an uplink slot, it corresponds to the first DMRS of the slot. Using LCP method 1, we get an estimate of the channel in Special slot Sim as ĝ_i,tm,s,awhich is a linear combination of gi,tm-5,s,a, . . . ,gi,tm-5-5*(ot-1),s,a based on the channel predictor coefficients for the ith intra-site interfering user. This estimate is used as the estimate of the channel in the first DMRS of Uplink slot U_tm+i, i.e., ĝ_j,tm+i,s,a=ĝ_,j,tm,s,a..

Let us concatenate across all antennas as

$\begin{array}{cc}{\hat{g}}_{i,t_m,s}=\left[\begin{array}{c}{\hat{g}}_{i,t_m,s,1}\\ ⋮\\ {\hat{g}}_{i,t_m,s,N_R}\end{array}\right]& \left(26\right)\end{array}$

The estimate of ICV based on the predicted intra-site interfering channels in the SRS symbol of Special slot S_tm is given as

$\begin{array}{cc}{\hat{C}}_{7.2}={\sigma }^{2}I+\frac{1}{12}\sum _{i=\text{?}}^{N_{u_{\text{?}}}}\sum _{\text{?}}^{}{\hat{g}}_{i,t_m,s}{\hat{g}}_{i,t_{m,s}}^H& \left(27\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

Using LCP method 2, we get an estimate of the channel in uplink Slot U_tm+i as ĝ_j,tm+i,s,a which is a linear combination of g_j,tm−5,s,a, . . . ,g_{j,tm−5−5·(ot−1),s,a} based on the channel predictor coefficients for thejth intra-site interfering user. Let us concatenate ĝ_j,tm+i,s,a across all antennas as

$\begin{array}{cc}{\hat{g}}_{j,t_m+1,s}=\left[\begin{array}{c}{\hat{g}}_{j,t_m+i,s,1}\\ ⋮\\ {\hat{g}}_{j,t_m+i,s,N_U}\end{array}\right]& \left(28\right)\end{array}$

The estimate of ICV based on the predicted intra-site interfering channels in the first DMRS in Uplink slot U_tm+1 is given as

$\begin{array}{cc}{\hat{C}}_{7.2}={\sigma }^{2}I+\frac{1}{12}\sum _{j=\text{?}}^{N_{u_{\text{?}}}}\sum _{\text{?}}^{}{\hat{g}}_{j,t_m+i,s}{\hat{g}}_{j,t_{\text{?}}}^H& \left(29\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

An estimate of ICV based on SRS symbol in Special slot S,,,,-s is given as

$\begin{array}{cc}{\hat{C}}_{7.2}={\sigma }^{2}I+\frac{1}{12}\sum _{i=1}^{N_{u_{\text{?}}}}\sum _{s\in r}^{}{g}_{i,t_m-5,s}{g}_{i,t_m-5,s}^H& \left(30\right)\end{array}$ $\mathrm{where}$ $\begin{array}{cc}{g}_{i,t_m-5,i}=\left[\begin{array}{c}{g}_{i,t_m-5,s,\text{?}}\\ ⋮\\ {\hat{g}}_{i,t_m-5,s,N_R}\end{array}\right]& \left(31\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

The various combining matrices for Split 7.2 are given as follows:

• • 1) MMSE-CM (7.2): The combining matrix W_RU=(HsasC iHs_Rs+I)⁻¹H_SRS^HĈ_7.2⁻¹, where Ĉ_7.2is based on equation (30) and W_DU=I. or W_DU as per equation (3). Alternatively, W_RU=H_SRS^H(H_SRSH_SRS^H+Ĉ_7.2)⁻¹
  • 2) CCM (7.2): The combining matrix W_RU=H_SRS^HĈ_7.2⁻¹, where Ĉ_7.2 is based on equation (30) and W_DU is based on equation (3).
  • 3) MMSE-CM-LCP (7.2): The combining matrix W_RU=(Ĥ_SRS^HĈ_7.2⁻¹Ĥ_DMRS+I)⁻¹Ĥ_DMRS^HC_7.2⁻¹, where Ĥ_DMRS ^H and Ĉ_7.2 are based on equation (14) and equation (27), respectively, for LCP Method 1. For LCP method 2, Ĥ_DMRS ^H and C_7.2 are based on equation (20) and equation (29), respectively. Furthermore, W_DU=I. or W_DU as per equation (3). Alternatively, W_RU=Ĥ_DMRS^H(Ĥ_DMRSĤ_DMRS^H+Ĉ_7.2)⁻¹
  • 4) CCM-LCP (7.2): The combining matrix W_RU=H_SRS^HĈ_7.2⁻¹, where Ĥ_DMRs and C_7.2 are based on equation (14) and equation (27), respectively, for LCP Method 1. For LCP method 2, Ĥ_DMRS^H and Ĉ_7.2 are based on equation (20) and equation (29), respectively. Furthermore, W_DU is computed as per equation (3).

Some points to be noted regarding the ICV for Split 7.2 and Split 7.3 include the following:

• • 1) The ICV for Split 7.3 in (23) depends on both the intra-site interfering channels (G) and the inter-site interfering channels (J). In functional Split 7.2, the DU has access to only intra-site interferers and not inter-site interferers. Consequently, the ICV for Split 7.2 (using equations (27), (29) and (30)) have contributions from only intra-site interferers but not inter-site interferers. As will be seen in the following sections, the performance of Split 7.2 (with linear channel prediction) can equal the performance of Split 7.3 only in the absence of inter-site interference. When inter-site interference is present, the performance of Split 7.2 consistently falls short of the performance of Split 7.3.

2) Note that ICV for Split 7.3 implicitly uses interfering channels at all subcarriers s in a RB r (23). It is for this reason that we have to use linear channel prediction at all subcarriers of intra-site users, as per equations (27), (29) and (30). It might appear that complexity of implementing training and prediction of channels of all intra-site interferers at all subcarriers of a RB is quite costly. However, as will be seen in the following sections, the linear channel predictor coefficients are substantially the same across RBs and slots. Therefore, we need not train the LCP coefficients for all subcarriers in an RB. Instead, we can just train the LCP for some subcarriers, and once trained, the same coefficients can be used in prediction across many RBs.

3) For Split 7.2, we compute the ICV based on LCP-estimated channel of intra-site interferers across all subcarriers in an RB, which needs to be done to match the performance of Split 7.3. The channel at all subcarriers can be calculated by taking the FFT of the received signal divided by the reference signal across subcarriers, separating the various FFT bins of users, and then taking the IFFT of the separated FFT bins.

4) Low complexity implementation of ICV for Split 7.2 is as per equation (25), which implements an approximate ICV (ICVA) based on an average channel across an RB of intra-site interferers.

In this section, some example applications of LCP in the context of Split 7.2 are discussed, as well as some low-complexity implementation alternatives of the techniques discussed above.

Downlink Precoding:

The first example application is downlink precoding. Consider a TDD system with uplink and downlink channel reciprocity. In downlink MU-MIMO, precoding is often employed at the base station (transmitter) to reduce interference in the downlink among many users. As an example, consider a pattern of slots DDDSUDDDSU and Downlink slot Dm., where the nearest Special slot is S_t.. The input-output equation at any OFDM data symbol in this slot D,.,, subcarrier s and RB r is given by

y=HPx+n  (32)

where x is N_D ×1 vector of downlink data to ND users, P is ND xND precoder a t transmitter of base station, H is the ND xND channel matrix from base station to all ND UEs, n is AWGN, y is N_D ×1 vector of received signals at NDusers stacked on top of one another. If we select the precoding matrix as P=H^H(HH)⁻¹, then the above equation reduces to y=x+n, which essentially means no inter-user interference in the downlink among the UEs. However, the base station (e.g., RU) needs to know the channel H in Downlink slot Dm+,. The DU has channel estimates at the SRSs of special slots S_tm,S_tm−5, . . . , and using these estimates across users and antennas for the given Subcarrier s in the RB r and LCP Method 2, the DU can predict H and send the precoder P to RU well ahead of the transmission of Downlink slot D_tm+i. Note that LCP Method 1 may not work well if the Downlink Slot D_tm+i is quite away from SRS in Special slot S_tm(especially for i =2,3, . . . ), in which case it becomes necessary to use LCP Method 2.

Low-Complexity Implementations:

• • 1) Adaptive Filters: The LCP coefficient vector p_t(as discussed above in the section discussing LCP methods for Split 7.2 in the absence of ICI) is learnt via computing an inverse of a o_t ×o_t matrix. An alternative is to use adaptive filters to learn p. (which does not involve any o_t×o_t matrix inversion). Adaptive filters, e.g., normalized least mean square (NLMS), affine projection (AP) and recursive least squares (RLS) can be used. While the RLS converges quickly compared to NLMS, it is computationally expensive. AP converges as good as RLS at a complexity of NLMS. Adaptive filters can be used across antennas for learning as discussed above in connection with LCP methods for Split 7.2 in the absence of ICI. Learning via adaptive filters needs to be careful as the filter coefficients can diverge in deep fades. Generally, a step-size controller can be employed, which freezes learning in deep fades, employs bigger step size initially so that the filter converges quickly, and in steady state employs a smaller step size so that steady-state error is small.

2) Kalman Filters: Kalman filters are used to study channel estimation. A Kalman filter inherently has a LCP coefficient vector to govern the state equation. As will be seen in the sections below, performance of Split 7.2 is substantially the same as Split 7.3 when LCP prediction order o_t=7 (though lower orders such as o_t=3 also result in good performances). A Kalman filter performs prediction and correction based on noisy observations, while an LCP does only prediction. A Kalman filter with low order o_t=1,2 as part of its state equation may achieve performances substantially the same as LCP of order o_t=7. In that case, the complexity comparisons between the Kalman filter with o_t=1,2 and LCP with o_t=7 need to be consider.

3) Fractional Delay Filters: In LCP Method 2, we used Whittaker-Shannon to reconstruct the channel between two SRSs in special slots. This used both future and past SRSs for reconstruction. The number of SRS symbols used on either side of reconstruction can be reduced if one uses fractional delay infinite impulse response (IIR) filters.

4) ICV Updates: The error in the estimate of ICV depends on the prediction order ot. If the prediction order is less (e.g., for complexity reasons), the estimate of ICV has more error, but this error can be accounted for by using the prediction error variance during training of all users. We now describe how it is done. Assume training at Uplink slot U_tm+i , (this is where the channel is reconstructed in LCP Method 2) and nearest SRS in Special slot is S_tm. The actual channel of the jth intra-site interferer in Subcarrier s belonging to RB r and Antenna a is denoted, as before, by g_j,tm+i,s,a. For the sake of simplicity, let us assume a low prediction order o_t=2. The estimated channel of the jth intra-site interferer in Subcarrier s belonging to RB r and Antenna a is denoted by ĝ_j,tm+i,s,a. Note that ĝ_j,tm+i,s,a is a linear combination of g_j,tm−5,s,a and g_{j,tm−10,s,a} (SRS periodicity is five slots) and error in prediction is denoted as e_j,m+s,a=g_{j,tm +i,s,a}−ĝ_j,tm+i,s,a. Just like equation (28), we concatenate all g_j,tm+i,s,a across antennas to give the vector g_j,tm+i,s. Similarly, we concatenate e_j,tm+i,s,a into e_j,tm+i,s and ĝ_j,tm+i,s,a into ĝ_j,tm+i,s. We have

gj,tm+_i,s=ĝ_j,tm+i,s +e_j,tm+i,s  (33)

From the definition of LCP, error is orthogonal to input data, i.e., E{e,_j,tm+i,s,aĝ_j,tm+i,s,a*}=0. Furthermore, the channel across antennas is uncorrelated, implying E{e,_j,tm+i,s,aĝ_j,tm+i,s,a*}=0, a≠b. From this and equation (33), it follows

E{g,_j,tm+i,sg_j,tm+i,s^H}=E{g,_j,tm+i,sg_j,tm+i,s^H}+σ_i²I  (34)

where E{e_j,tm+i,s,ae_j,tm+i,s,a^H}=σ_i². An estimate of σ₁² is computed as

$\begin{array}{cc}{\hat{\sigma }}_i^2=\frac{1}{12N_R}\sum _{s\in r}{e}_{j,t_m+i,s}^H{e}_{j,t_m+i,s}=\frac{1}{12N_R}\sum _{s\in r}{\left(g_{j,t_m+i,s}-{\hat{g}}_{j,t_m+i,s}\right)}^H\left(g_{j,t_m+i,s}-{\hat{g}}_{j,t_m+i,s}\right)& \left(35\right)\end{array}$

which is computable as we are in training mode and have access to both g_j,m+i,s, ĝ_j,m+i,s. The ICV for Split 7.2 based on ĝ_j,m+i,s is given by equation (29) while the ICV for Split 7.2 based on actual channel value g_j,m+i,s is

$\begin{array}{cc}{\hat{C}}_{7.2}={\sigma }^{2}I+\frac{1}{12}\sum _{j=1}^{N_{u_{\text{?}}}}\sum _{s\in r}^{}{g}_{j,t_m+i,s}{g}_{j,t_m+i,s}^H.& \left(36\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

So, it follows from equation (34) that the updated version of equation (29) is

$\begin{array}{cc}{\hat{C}}_{7.2}={\sigma }^{2}I+\frac{1}{12}\sum _{j=1}^{N_{u_{\text{?}}}}\sum _{s\in r}^{}{g}_{j,t_m+i,s}{g}_{j,t_m+i,s}^H+\sum _j^{N_{u_{\text{?}}}}{\sigma }_i^{2}I.& \left(37\right)\end{array}$ $\text{?}\text{indicates text missing or illegible when filed}$

5) LCP Coefficients: LCP coefficient vector p. depends on the channeP-³⁸,E variation across time. This in turn is dependent on delays, scattering, angle of arrival at base station, speed etc. To a large extent, it can be dependent on speed of the UE. In that case, we can find speed of the UE by computing power spectral density of SRS symbols and then using a look-up table to determine the LCP coefficient of a given order. We then do not have to spend resources in learning, and we can just use the coefficients directly in prediction.

6) Training in FFT Domain: Due to frequency selectivity, channel across subcarriers varies slowly. If we take a 12-point FFT of the channel in an RB, it can be accurately represented by a few FFT bins only, e.g., FFT bins 1, 2, 3, 11, and 12 (bin index starts at 1). If frequency-selectivity is less, it can even be represented by FFT bins 1, 2, and 12, for example. Now if we take the IFFT of these FFT bins (other bins being zero), the channel across subcarriers in the RB can be represented quite accurately with around 5% error. In general, for multi-RB, we can take FFT of all subcarriers across the RBs, and use only very few FFT bins. To get subcarrier-by-subcarrier channel and separate the multi-users, we need to get into FFT domain, so we can learn on the selected reduced FFT bins and then reconstruct the channels at subcarriers over the RBs.

SRS Channel Estimation: In an example method, we take FFT of raw channel across subcarriers, drop some FFT bins which correspond to noise and retain only the low-pass FFT bins corresponding to the channel, and then generate IFFT. This is called denoising, which works because the channel varies as a low-pass signal across subcarriers. We can extend this to two dimensions for SRS, i.e., time and frequency. The channel across various SRS also varies slowly (depends on Doppler and a low-pass signal). So, if we take a 2D-FFT of raw SRS channel (received signal divided by reference sequence), we can separate the various users in 2D-FFT domain, remove the noise 2D-FFT bins, and then generate IFFT to obtain SRS channel estimations across various SRS OFDM symbols/RBs for each user.

Delay-Doppler-based Channel Prediction and SRS Capacity Improvement:

Delay-Doppler-based channel prediction is implemented when we have modeled all the multipath channel taps along with delays. These channel taps can be tracked and predicted like the channel prediction described in the present disclosure. The key aspect is at any time a set of frequency subcarriers can be converted into time-domain multipath channel taps (the channel taps of many users derived from the same subcarriers can be separated.). These channel taps are computed for every SRS symbol and learning/prediction can happen for each channel tap. The Doppler associated with each channel tap can vary, but we can assume the Doppler to be constant for a few taps and learning can happen for selected few channel taps only, and the prediction be used on all channel taps (based on learning of some selected taps). Once the future channel taps are predicted/estimated, we can reconstruct/predict the channel in any RB. In present disclosure, we use full-band scheduling. However, with the proposed usage of linear prediction in time domain of channel taps, sub-band scheduling can be employed as one can estimate/update the channel taps in time domain from any sub-band. This will increase the SRS capacity, but is computationally expensive.

SIMULATION RESULTS

The channel model used in these simulations corresponds to CDL-B (see, e.g., 3GPP TR 38.901, “Study on Channel Model for Frequencies from 0.5 to 100 GHz,” 3GPP, V14.3.0, Dec. 2017). Furthermore, we assume that v=30 km/hr,fc=3.5 GHz, N_R =64, N_U=4, delay spread (ds) is 1000 ns. The delay spread corresponds to the highest value in Table 7.7.3-1 in 3GPP TR 38.901, “Study on Channel Model for Frequencies from 0.5 to 100 GHz,” 3GPP, V14.3.0, Dec. 2017. This means the frequency-selectivity is the highest and the variation of HDMRsw.r.t H in equation (1) is the largest. Furthermore, we assume perfect channel estimation, i.e., we have perfect knowledge of Hi, H in equation (1). We assume a 4×8×2 panel of antennas with four in vertical direction, eight in horizontal direction and two cross polarized antennas in each location. We consider the DSUUU pattern, 30 kHz subcarrier spacing, where each slot is 0.5 ms in duration and TsRs=25 ms (SRS periodicity of five slots).

FIG. 4 illustrates the effect of using predictions of both LCP methods vs. not using the LCP methods, which FIG. 4 uses plots of the normalized error square (y-axis labeled as NES) vs. prediction order o_t (on the x-axis). Assume we are detecting an uplink Slot U_tm+i ,, and the nearest preceding special slot is Srm. The prediction error is defined as the difference in the actual channel ω_u,tm+i,r,a value and predicted value. The predicted values are p_u,tm+i,r,a=ω_u,tm+i,r,a and p_u,tm+i,r,a=θ_u,tm,r,a, for LCP Method 1 and 2, respectively. The SRS error is defined as the difference in the actual channel value and value of the channel at S_tm−5. The normalized prediction error square (NPES) is defined as the ratio of expectation of the square of the absolute value of the prediction error to the expectation of the square of the absolute channel value. The expectation is with respect to users (all users have same speed and hence we assume the same random process statistically), antennas, RBs and across time slots. The normalized SRS error square (NSES) is defined as the ratio of expectation of the square of the absolute value of the SRS error to the expectation of the square of the absolute channel value. The resulting equations for NPES and NSES are as follows:

$\begin{array}{c}\mathrm{NPES}=\frac{\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}{❘\text{?}❘}^2}{\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}{❘\text{?}❘}^2}\\ \mathrm{NSES}=\frac{\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}{❘\text{?}❘}^2}{\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}\sum _{\text{?}}{❘\text{?}❘}^2}\end{array}.$ $\text{?}\text{indicates text missing or illegible when filed}$

Note that, but for the LCP methods, we will have to use the channel estimate of the of the SRS at Sm,-s. As shown in FIG. 4, the NSES is quite high compared to NPES, and it is evident the LCP methods improve the performance by reducing the NPES relative to the NSES. NPES for LCP method 2 is worse than that of LCP method 1 as there is reconstruction error associated with the learning.

FIG. 5 depicts the normalized reconstruction error using Whittaker-Shanon interpolation and Nyquist-Shannon sampling theorem, e.g., for LCP Method 2. We consider 771 OFDM symbols or 56 slots (0.0275 s). Beginning from the first OFDM symbol, every 70th OFDM symbol is an SRS symbol that is known. Using only the channel estimates (perfect) of the SRS symbols the channel estimates of all the OFDM symbols are reconstructed using Whittaker-Shanon interpolation and Nyquist-Shannon sampling theorem. The channel estimates are denoted by x(n,i,a),n=0, . . . ,771, where i is the iteration index, a is the antenna index and n is the OFDM symbol index. For the sake of simplicity, we assume only one user. The reconstructed channel estimates are denoted by x∧(n,i,a),n=0, . . . ,771. The normalized reconstruction error square is defined as |x∧(n,i,a)−x(n,i,a)|² ETF X(n. i., a?-TT a)²

$\begin{array}{cc}\mathrm{NRES}\left(n\right)=\frac{\sum _{\alpha }\sum _i{❘\hat{x}\left(n,i,a\right)-x\left(n,i,a\right)❘}^2}{\sum _i{❘x\left(n,i,a\right)❘}^2}.& \left(39\right)\end{array}$

The channel estimate at the middle is as small as 0.5%. Each iteration is over an RB (one channel estimate per RB) and 56 slots. A time-frequency observation of 50 RBs and 105 slots was used. A sliding window of 56 slots was used as we move from one iteration to another in the time domain.

FIGS. 6 and 8 depict performance comparisons of various methods discussed in the present disclosure. The simulation corresponds to 2000 iterations. The dataset for each iteration is from one RB over 13 special slots (SRS symbols) followed by three uplink slots (the three uplink slots are just after the 13th special slot). The datasets for iterations first vary over a channel bandwidth of 50 RBs while the time domain slots are held constant and then a sliding window that slides by five slots (SRS periodicity) is used to move in the time domain to get the next 50 datasets for iterations over the channel bandwidth of 50 RBs. We use prediction order o_t=7. In any dataset (iteration) that requires learning, the first eight special slots (SRS) are used for training. Learning is over only one RB and one time instant, i.e., r₁=r_A, t₁ =t_n . For every 400 iterations, we use LCP coefficients learnt in the first of the 400 iterations. This is ideal from a complexity and scheduling viewpoint as we can learn in one RB and time slot and use it to predict in another RB and time slot. At the time of learning, we do not know where (in which RB) the user will be scheduled in future. In each dataset, the channel estimates of the first 12 SRS symbols in the RB is used to reconstruct the channel estimate of all OFDM symbols between the first SRS symbol and 12th SRS symbol (a region of 56 slots or 770 OFDM symbols). The prediction in any dataset (iteration) uses channel estimates of SRS symbols from special slots six to 12, to predict the channel estimate in the SRS symbol of the 13th special slot (LP Method 1) or the channel estimate of the first DMRS of any of the three uplink slots following the 13th special slot (LP method 2). Mathematically, the dataset for the kth iteration is taken from RB r=k mod 50 and Slots

$S_{t+5\lfloor \frac{k}{50}\rfloor },S_{t+5\lfloor \frac{k}{50}\rfloor +5},\dots ,S_{t+5\lfloor \frac{k}{50}\rfloor +55},S_{t+5\lfloor \frac{k}{50}\rfloor +60},$ $U_{t+5\lfloor \frac{k}{50}\rfloor +61},U_{t+5\lfloor \frac{k}{50}\rfloor +62},U_{t+5\lfloor \frac{k}{50}\rfloor +63}.$

In FIG. 6, which corresponds to only FUS, we can see that the MMSE-CM 7.2 method has the worst performance as it is entirely based on a 2.5 ms (or five slots) old SRS value. The CCM 7.2 method has improved performance compared to MMSE-CM 7.2 method as it is partly based on an old SRS channel estimate (only W_Ru part is based on the 2.5 ms (or five slots) old SRS value, while the W_DU part is based on the effective channel matrix H DMRS corresponding to the first DMRS of the uplink slot being decoded). Both methods, however, have a big performance gap relative to the Split 7.3 method. The LCP Method 1 for prediction order of seven exactly matches the performance of Split 7.3 method (i.e., the upper bound) for FUS, thereby overcoming the channel aging and fronthaul constraints. This matching of the upper bound of Split 7.3 method via LCP is one of the goals of the present disclosure. FIG. 6 also shows LCP Method 1 performances for prediction orders o_t=2, 4, and these are close to the upper bound of the Split 7.3 method, and much better than MMSE-CM 7.2 and CCM 7.2, thereby emphasizing the importance of the LCP methods.

FIG. 8 illustrates the comparisons between LCP Methods 1 and 2 for FUS and TUS. As discussed previously, LCP Method 1 (Split 7.2) works well for FUS and attains the same upper bound performance of Split 7.3 method. However, LCP method 1 will exhibit loss of performance for TUS, while LCP Method 2 attains the same performance as Split 7.3 method, as shown in FIG. 8.

Ideally, one can learn the LCP coefficients across time, RBs and antennas as given in equation (12). The channel is a random process characterized by Doppler spread f_D which is dependent on the speed of the UE v and carrier frequency fc. The channels at various RBs and antennas can be considered to be realizations of the same random process. Doppler spread which is dependent on speed and carrier frequency is expected to change little across the RBs, as change in frequency from RB to RB is very little compared to carrier frequency. Consequently, we assume that LCP coefficients are constant across RBs, antennas and time, as shown in FIG. 7.

For O-RAN massive MIMO systems, there is a loss of performance of Split 7.2 relative to Split 7.3 for mobility use cases in the uplink when conventional methods are used. The two example linear channel prediction methods provided in the present disclosure overcome this loss of performance of Split 7.2 and help the Split 7.2 systems achieve substantially the same performance as Split 7.3 systems. Results for four uplink users at speeds of 30 km/hr at a carrier frequency of 3.5 GHz showed that Split 7.2 achieved the same performance as Split 7.3 systems, as illustrated in FIG. 7. FIG. 7 additionally illustrates that LCP Method 2 overcomes the limitations of LCP Method 1 for TUS.

In summary, several example embodiments of the method according to the present disclosure are listed below:

1) Learning and predicting channel at each subcarrier in a future SRS symbol from present and past channel values at subcarriers of past and present SRS symbols for each user (desired and intra-site) using LP Method 1.

2) Learning and predicting average of channels across subcarriers in a RB for a future SRS symbol from present and past such average values corresponding to past and present SRS symbols for each user (desired and intra) using LP Method 1.

3) Learning and predicting channel at each subcarrier in a future PUSCH symbol from present and past channel values at subcarriers of past and present SRS symbols for each user (desired and intra) using LP Method 2.

4) Learning and predicting average of channels across subcarriers in a RB for a future PUSCH symbol from present and past such average values corresponding to past and present SRS symbols for each user (desired and intra) using LP Method 2.

5) Using the Method of 1) above to compute ICV.

6) Using the Method of 2) above to compute ICV.

7) Using the Method of 3) above to compute ICV.

8) Using the Method of 4) above to compute ICV.

9) Using Method of 2) above to compute desired Channel matrix.

10) Using Method of 4) above to compute desired Channel matrix.

11) Using any of Methods 5)-8) for ICV in conjunction with any of Methods 9)-10) for channel matrix of desired users to decode desired users as per equation (24).

12) For determining LCP coefficients of any user, learning can be done at only one subcarrier and this can be used for prediction across RBs and time domain slots.

13) Use of LCP methods for downlink precoding for MU-MIMO.

14) The use of adaptive filters to determine LCP coefficients.

15) The use of Kalman filters to determine LCP coefficients.

16) The use of fractional delay IIR filters for reconstruction in LP Method 2.

17) The use of Nyquist sampling for reconstruction in LP Method 2.

18) Updating ICV corresponding to a lower prediction order to a more accurate version.

19) Determining speed of users and using a LUT to determine LCP coefficients.

20) Taking FFT of raw channel estimates across subcarriers, and using linear prediction only on a subset of FFT bins, predicting those FFT bins and in turn reconstructing predicted channels across subcarriers for many RBs and multiple users.

21) Going from frequency-domain to time domain and computing multipath tap channel values. Using LCP on the time domain taps and predicting future time domain taps, thereby being in a position to compute the channel for any RB in a future time domain slot. Using this to improve SRS capacity via sub-band scheduling.

22) Using 2D-FFT for SRS channel estimation.

23) A method of linear channel prediction (LCP) in an Open Radio Access Network (O-RAN) massive Multiple Input Multiple Output (MIMO) system, comprising: predicting a channel corresponding to a sounding reference signal (SRS) symbol closest to a target uplink (UL) slot to be decoded, wherein said predicting is performed based on at least one previous SRS symbol; and using the predicted channel as a combining matrix to be applied in a radio unit (RU) for decoding the target UL slot.

24) A method of linear channel prediction (LCP) in an Open Radio Access Network (O-RAN) massive Multiple Input Multiple Output (MIMO) system, comprising: i) reconstructing a channel between the two periodic SRS symbols using past and future SRS symbols; ii) training for predicting a channel between two periodic SRS symbols; and iii) using the predicted channel corresponding to the UL slot located between the slots containing the two periodic SRS symbols and decoding the UL slot using a combining matrix based on the predicted channel.

Definitions

• 3GPP 3rd Generation Partnership Project
• 5G N_R TDD 5G New Radio Time-Division Duplex.
• AWGN Additive White Gaussian Noise
• ADC Analog to Digital Converter
• BS Base Station
• CE Channel Estimation
• CP Cyclic Prefix
• CSF Current Subframe
• CSI Channel State Information
• DL Downlink
• DMRS Demodulation Reference Signals
• DU Distributed Unit
• eNB eNodeB
• FEC Forward Error Correction
• FFT Fast Fourier Transform
• FH Fronthaul
• GSMA Global System for Mobile Communications Association
• IFFT Inverse Fast Fourier Transform
• IQ In-phase/Quadrature-phase
• IRC Interference Rejection Combining
• LTE Long-Term Evolution
• LCP Linear Channel Prediction
• MAC Medium Access Layer
• MIMO Multiple Input Multiple Output
• mMIMO Massive MIMO
• MMSE Minimum Mean Squared Error
• MU-MIMO Multi-User MIMO
• NG-RAN Next Generation Radio Access Network
• OFDM Orthogonal Frequency Division Multiplexing
• O-DU Open Distributed Unit
• O-RAN Open Radio Access Network
• O-RU Open Radio Unit
• PUSCH Physical Uplink Shared Channel
• RAN Radio Access Network
• RB Resource Block
• RF Radio Frequency
• SF Subframe
• SFN System Frame Number
• SRS Sounding Reference Signals
• SSF Special Subframe
• TDD Time division duplex
• UE User Equipment
• UL Uplink

Claims

1. A method of linear channel prediction (LCP) in an Open Radio Access Network (0-RAN) massive Multiple Input Multiple Output (MIMO) system, comprising: predicting a channel associated with a target uplink (UL) slot to be decoded, wherein said predicting is performed based on at least one previous SRS symbol; andusing the predicted channel as a basis for a combining matrix to be applied in a radio unit (RU) for decoding the target UL slot.

2. The method according to claim 1, wherein the predicting comprises: predicting a channel corresponding to a sounding reference signal (SRS) symbol closest to a target uplink (UL) slot to be decoded.

3. The method according to claim 2, wherein the predicted channel is used as the combining matrix to be applied in the RU for decoding the target UL slot.

4. The method according to claim 3, wherein the predicting comprises: training with a latest SRS symbol as the desired channel and past SRS symbols as inputs to a prediction module.

5. The method according to claim 4, wherein prediction coefficients are obtained by the training.

6. The method according to claim 5, wherein the predicting further comprises: using the prediction coefficients to predict at least one future SRS symbol based on the at least one previous SRS symbol.

7. The method according to claim 6, wherein the predicted SRS symbol is used to derive the combining matrix to be applied in the RU for decoding the target UL slot.

8. The method according to claim 1, wherein the predicting comprises: predicting a channel of the target uplink slot from at least previous SRS symbols.

9. The method according to claim 8, wherein the combining matrix is built based on the predicted channel of the target uplink slot.

10. The method according to claim 1, wherein the predicting comprises: training for predicting a channel between two periodic SRS symbols; anddecoding the target UL slot using a combining matrix based on a predicted channel corresponding to the target UL slot located between slots containing the two periodic SRS symbols.

11. The method according to claim 11, further comprising: prior to the training for predicting, sampling a channel between two periodic SRS symbols for reconstructing the channel between two periodic SRS symbols.

12. The method according to claim 11, wherein the sampling the channel is performed at a frequency greater than the Nyquist sampling frequency.

13. The method according to claim 11, wherein prediction coefficients are obtained by the training.

14. The method according to claim 13, wherein the prediction coefficients are learned by an adaptive filter configured as one of a normalized least mean square (NLMS) filter, affine projection (AP) filter, and recursive least squares (RLS) filter.

15. The method according to claim 13, wherein the prediction coefficients are learned by a Kalman filter.

16. The method according to claim 11, wherein fractional delay infinite impulse response (IR) filter is used for reconstructing the channel between two periodic SRS symbols.

17. The method according to claim 5, wherein the prediction coefficients are learned by an adaptive filter configured as one of a normalized least mean square (NLMS) filter, affine projection (AP) filter, and recursive least squares (RLS) filter.

18. The method according to claim 5, wherein the prediction coefficients are learned by a Kalman filter.

19. The method according to claim 2, wherein the predicting comprises: training with a latest SRS symbol as the desired channel and past SRS symbols as inputs to a prediction module.

20. The method according to claim 19, wherein prediction coefficients are obtained by the training, and the prediction coefficients are used to predict at least one future SRS symbol based on the at least one previous SRS symbol.