Design of Full-Duplex Millimeter-Wave Integrated Access and Backhaul Networks
Junkai Zhang, Navneet Garg, Mark Holm, Tharmalingam Ratnarajah
11 Design of Full-Duplex Millimeter-Wave IntegratedAccess and Backhaul Networks
Junkai Zhang,
Student Member, IEEE , Navneet Garg,
Member, IEEE,
Mark Holm,
Member, IEEE, and Tharmalingam Ratnarajah,
Senior Member, IEEE A BSTRACT
One of the key technologies for the future cellular networksis full-duplex (FD) enabled Integrated Access and Backhaul(IAB) networks operating in the millimeter-wave (mmWave)frequencies. The main challenge in realizing the FD-IABnetworks is mitigating the impact of self-interference (SI) inthe wideband mmWave frequencies. In this article, we firstintroduce the 3GPP IAB network architectures and widebandmmWave channel models. By utilizing the subarray-basedhybrid precoding scheme, at the FD-IAB-node, multiuserinterference is mitigated using zero-forcing (ZF) at the trans-mitter, whereas the residual SI after successfully deployingantenna and analog cancellation is canceled by minimummean square error (MMSE) baseband combiner at the receiver.The spectral efficiency (SE) is evaluated for the RF insertionloss (RFIL) with different kinds of phase shifters and thechannel uncertainty. Simulation results show that, in thepresence of the RFIL, the almost double SE, which is closeto that obtained from fully connected hybrid precoding, canbe achieved as compared with half-duplex systems, when theuncertainties are of low strength.I
NTRODUCTION
The key technologies, namely, millimeter-wave (mmWave)wideband communications, full-duplex (FD) transmissions,and Integrated Access and Backhaul (IAB) networks, areemerging as the backbone of 5G and beyond communications.A large bandwidth provided by mmWave systems can beexploited for wideband transmissions to increase data rates,which are orders of magnitude more than that of the currentmicrowave systems. However, a beamformed array with alarge number of antennas is needed to compensate for thehigher path loss at mmWave frequencies [1]. Moreover, toenhance the coverage, dense deployment of multi-antennaaccess points has been considered as a promising approach.However, providing traditional fiber backhauling connectionto all these small cells is not possible both economically and
This work was accepted for publication in the special issue of Full DuplexCommunications Theory, Standardization and Practice, in IEEE WirelessCommunications, February, 2021. Manuscript received May 13, 2020; revisedSeptember 3, 2020; accepted October 19, 2020. Date of current versionDecember 15, 2020.J. Zhang, N. Garg, and T. Ratnarajah are with Institute for DigitalCommunications, The University of Edinburgh, Edinburgh, EH9 3FG, UK(e-mail: { jzhang15, ngarg, T. Ratnarajah } @ed.ac.uk.M. Holm is with Radio Basestation Systems Department,Huawei Technologies (Sweden) AB, Gothenburg, Sweden. (e-mail:[email protected]). physically. To address this issue, 3GPP proposed cost-effectivedense deployment of wireless backhauling through IAB-nodesto achieve promising gains even under higher mobile datatraffic [2].Moreover, to leverage the full benefits of IAB networkswith the mmWave wideband, the IAB-nodes are set to oper-ate in the FD mode. Compared with the half-duplex (HD)transmission, FD can enhance the spectral efficiency (SE)and reduce the communication delay without any requirementfor the guard time/band [3]. Unlike traditional microwavecommunications, where full digital baseband (BB) precodingschemes are sufficient, the hybrid precoding is essential inmmWave communications [1]. For wideband mmWave-FD-IAB networks, the hardware efficient subarray based hybridprecoding is adopted in this article.Since in a FD-IAB network, the access and the backhaulcommunications occur at the same time-frequency resource,it naturally gives rise to self-interference (SI) at the receiverof the FD-IAB-node. Typically, the magnitude of the SI canbe more than 100 dB, which is much stronger than thesignal of interest, as studied in [4]. Such a high SI powercan significantly exceed the hardware dynamic range anddistort the benefits of FD transmission. Thus, it is impor-tant to reduce SI power before the down-conversion. In themicrowave communications, successful SI cancellation (SIC)can be achieved at the antenna domain (i.e., by deployingspecial antenna isolation), the RF domain (i.e., by replicatingthe SI channel and subtracting it from the received signal),and the digital domain (i.e., by canceling the residual SI(RSI) after RF cancellation by beamformer design). Usually,a combination of these stages has shown satisfactory results[3], which we also expect to provide a good solution forthe mmWave wideband communications. In this article, wemainly focus on the design of the digital cancellation, wherethe antenna isolation and the RF cancellation are assumed tobe successfully achieved. Therefore, only the RSI signal willbe handled in the digital domain.In this article, we first introduce the fundamental 3GPPnetwork architectures for the FD-IAB systems, followed bythe description of the general mmWave and the SI channelmodels. Next, a hybrid analog/digital transceiver design viathe cost-efficient subarray structure for the multiuser scenariois explained. The multiuser interference (MUI) at the trans-mitter of the IAB-node and the RSI at the receiver of the IAB-node are mitigated by zero-forcing (ZF) and minimum meansquared error (MMSE) in the digital BB domain, respectively.Further, the performance limitations of FD enabled multiuser a r X i v : . [ c s . I T ] J a n GC IAB donorgNB IAB-node UEWireline Wireless
NR Uu
Wireless
NR Uu (a)
NGC IAB donorgNB IAB-node UEWireline Wireless
NR Uu
Wireless
NR Uu
EPC eNB Wireless
LTE Uu (b)
IAB donorgNB IAB-node UEWireline Wireless
NR Uu
Wireless
NR Uu
EPC eNB Wireless
LTE Uu (c)Fig. 1. Examples for IAB-FD network architectures operating in SA modeand NSA mode: a) UE: SA with NGC, IAB-node: SA with NGC; b) UE:NSA with EPC, IAB-node: SA with NGC; c) UE: NSA with EPC, IAB-node:NSA with EPC. mmWave-IAB networks under subarray hybrid precodingstructure are studied in the presence of the RF insertionloss (RFIL) and the channel estimation error (CEE). Withthe RFIL, simulations show that the SE performance of thefully connected hybrid precoding structure is similar to thatfor the subarray-based hybrid precoding structure. Moreover,as the CEE increases, the rate improvement of FD over HDdecreases. Besides, the SE intersection point of FD and HDthat appears at the backhaul link enables the understandingof the maximum achievable digital cancellation, which willencourage the development of advanced hybrid transceiverswith efficient resource allocation schemes in the future.3GPP N
ETWORK A RCHITECTURES
The 3rd Generation Partnership Project (3GPP) release 16explores the standards for 5G new radio (NR) communica-tions. IAB architectures, radio protocols, and physical layeraspects related to relaying of access traffic by sharing radioresources between access and backhaul links are investigated in the technical specification TR 38.874 [5]. These initialstudies show the benefits of in-band backhauling over out-of-band backhauling for access links. However, these fundamen-tal results for FD operations are still in its infancy. Further,the knowledge of the impact of FD operations at mmWavefrequencies is also limited, since the wideband channel modelfor FD operations still needs thorough investigation. Accord-ing to the 3GPP specification in [5], the IAB systems aretypically deployed in two modes, namely standalone (SA)mode, and non-standalone (NSA) mode, as shown in Fig. 1. Inthe SA mode shown in Fig. 1(a), the IAB-node connects to the5G next-generation core (NGC) network via the IAB donor(gNB), and the UE also operates in the SA mode (i.e., it onlyconnects to the IAB-node). In Fig. 1(b), the UE is connectedin the NSA manner, while the IAB-node is in the SA mode.In this scenario, both the Long Term Evolution (LTE) radioand the NR can be used for the UE, and NR links are utilizedfor backhauling. Further, if the IAB-node works in the NSAmode, it is also connected to the eNB nodes (i.e., the 4G basestations), as shown in Fig. 1(c). Thus, a UE in the NSA modecan choose to connect the IAB-connected-eNB or differentone. In the third scenario, the IAB-node can utilize the LTElinks for initial access, route selection, etc.A multihop mmWave IAB networks in SA mode is shownin Fig. 2(a). In this figure, there are three kinds of nodes listedas follows, • A single logical IAB donor, which is the source node,also known as the gNB. It takes the responsibility offunctionality and splits according to the 3GPP NG-RANarchitecture [6]. Usually, the gNB has a wired connectionto the core network (NGC) and has wireless connectionsto other nodes. • IAB-nodes, which wirelessly communicate with bothbackhaul and access links, provide FD operations, andperform IAB specific tasks such as resource allocation,route selection, and optimization. The IAB-nodes can beconnected to other HD-IAB-nodes or FD-IAB-nodes. • UE nodes, which request and receive the contents via FDor HD operation. Since UEs operate in the SA mode, theyonly connect to the IAB-nodes.Typically, the IAB-node enables not only UEs but also otherFD/HD-IAB-nodes to communicate with the gNB. In the SAarchitecture illustrated in Fig 2(a), IAB-nodes forward thebackhaul traffic to the core network in different spectrum,whereas with this general star topology, Taghizadeh et al .[7] consider a central station delivering the backhaul trafficfrom multiple nodes, which may require efficient interferencemanagement schemes.There are two kinds of topology models to characterizesuch multihop networks. The first one is the spanning tree(ST) model, where one IAB-node connects to only one parentnode (i.e., the IAB donor or another IAB-node). The secondmodel uses the directed acyclic graphs (DAG), where oneIAB-node has multiple parent nodes, or has multiple routesto one parent node, or a combination of these two cases. [5],[8]. These ST and DAG models for Fig. 2(a) are difficultto analyze from the physical layer perspective. Thus, for the ore Network IAB donorgNB
SI SISI
FDIAB-nodeFDIAB-node FDIAB-nodeUEUE SI FDIAB-nodeHDIAB-node UEWirelineWireless
Backhaul link
Wireless
Access link (a)
NGC IAB donorgNB IAB-nodeUENG IAB-nodeCU DU MT DU MT DUNR Uu NR Uu UE UENR Uu NR Uu NR UuRLC RLCF1*F1* (b)Fig. 2. a) Illustration of multihop mmWave-FD-IAB network architecturediagram in SA mode; b) CU/DU split architecture for multihop IAB system. multihop IAB networks, a simplified version, i.e., the CentralUnit (CU)/Distributed Unit (DU) split architecture is preferredin studies [2], [8], and is shown in Fig. 2(b), where theCU and the DU represent external interfaces of the node. Inthis architecture, the IAB-node has two NR functional units,viz., the Mobile Termination (MT) unit, which controls theupstream link connection with the IAB donor or the IAB-node;and the DU, which provides connections to UEs or MTs onother IAB-nodes of the downstream link. The IAB donor hastwo functional units as well, viz., the CU is responsible forserving the DUs on all IAB-nodes and the donor itself, whilethe DU provides support to the UEs and the MTs on all IAB-nodes. The F1* function connects the interface of the IAB-node to the interface of the IAB donor. It runs on the RadioLink Control (RLC) channels, representing the connectionsbetween the DU and the downlink MT or UEs.C
HANNEL M ODELS
General mmWave Channel
The mmWave channel has several characteristics that dif-ferentiate it from the traditional microwave channels, suchas higher path loss (owing to higher operating frequencies),the spatial selectivity (due to high path losses and beam-forming), and increased correlation among antennas (due todensely collocated arrays). These distinctive characteristicsimply that the statistical fading distributions such as theRayleigh distribution, used in traditional wireless channelsbecome inaccurate, since the number of fading paths is small.Hence, the mmWave channel between two different nodes islikely modeled as a geometric wideband frequency selective channel according to the extended Saleh-Valenzuela model,studied in [1], [9].An Orthogonal Frequency Division Multiplexing (OFDM)system with K subcarriers is adopted, where D cyclic prefix(CP) is added to avoid the Inter Symbol Interference (ISI).For each of the D taps of the wideband channel, scatterersin the area contribute to multiple propagation paths. Thesereflected multipath components (rays) arrive in clusters, whichcause the sparse nature in the channel response. The value ofthe d th tap of the channel is modeled using the product ofthe complex random gain, the complex exponential of anglesof arrival and departure (AoAs/AoDs), and the pulse-shapingfilter. The complex random gain of each ray has the magnitudefollowing the Rayleigh distribution with the parameter definedby the number of total paths. For the uniform planar arrays(UPAs), the central azimuth AoAs/AoDs of fading paths (rays)in each cluster are uniformly distributed in [ − π, π ] , and thecorresponding central elevation AoAs/AoDs are uniformlydistributed in [ − π/ , π/ . In each cluster, these azimuth andelevation angles of the rays are assumed to have the Laplaciandistribution with a given angle spread. The raised cosine pulseshaping filter is utilized with sampling time T s , evaluated at dT s − τ c,l seconds, where τ c,l is the path delay of the l th rayin the c th cluster and is uniformly distributed in [0 , DT s ] . Theclose-in (CI) path loss model with a reference distance of 1mis introduced to capture the average path loss. Ultimately, thechannel at subcarrier k = 1 , , ..., K is given by the discreteFourier transform (DFT) of the delay- d channel. Self-Interference Channel
The FD-IAB-node is comprised of a transmit antenna arrayand a receive antenna array. In FD operations, a mmWaveSI channel is defined as the mmWave channel between thetransmit antenna and the receiver antenna at the IAB-node.Through measurements, the mmWave SI channel is verified tohave both line-of-sight (LOS) and non-line-of-sight (NLOS)components [4]. The LOS component accounts for a deter-ministic direct path loss. Its strength is very high due to avery short distance between the transceiver of the IAB-nodeand is assumed to adopt a near-field model, since the distancebetween the transceiver is smaller than D /λ , where D isthe antenna aperture diameter, and λ is the wavelength [3].The coefficient of the LOS channel matrix depends on thedistance between the individual elements of the transceiver.The NLOS component indicates random components causedby reflections from obstacles around the IAB-node, where thegeneral mmWave channel model may be acceptable, exceptwith a smaller number of rays. A Rician-alike channel modelcould be utilized to model the SI channel due to a strong LOSpath. A detailed hypothetical wideband mmWave SI channelmodel is formulated in our recent work [10]. It is worth notingthat there is still ambiguity in characterizing the mmWave SIchannel model in the literature.A study in [3] shows that the resulting SI channel issparse and low rank. Unfortunately, as mentioned in [11],the difficulties of SIC arises due to its inability to cancelthe NLOS component of the SI signal by the three-stageIC scheme. It is due to the fact that the present SI channelestimation methods have proved to be inaccurate due to thestrong antenna correlation in the near-field region. Moreover,in general, the channel estimation for microwave communica-tions assumes the steady oscillator phase noise (PN), however,for mmWave communications, this assumption can cause largeestimation error, since the PN changes rapidly and cannotbe ignored. In [11], with the Rician SI channel model, ajoint SI channel and PN estimation algorithm for mmWavecommunications using the Kalman filter is proposed, which isshown to achieve its mean squared error (MSE) lower boundsuccessfully. With their efficient estimator, the RSI can bedecreased to an acceptable amount.The CEE, ∆ SI [ k ] , is introduced to model the imperfect RFeffective SI channel and analyze the corresponding systemperformance. The perfect RF effective SI channel (i.e., theproduct of the RF combiner, the SI channel matrix, and theRF precoder) at the k th subcarrier is assumed to be the sum ofthe estimated RF effective SI channel and the CEE. The CEEis assumed to be Gaussian with zero mean and small variance[12]. Note that the estimated effective channel after analogcancellation is used to design the precoders at transmittersand the combiners at receivers to cancel the remaining SI.However, interference leakage occurs due to the CEE andresults in the RSI power. The impact of the CEE on the systemcapacity is given in the later section.H YBRID T RANSCEIVER D ESIGN
Since the wideband channel is frequency selective, eachnode adopts an OFDM system, ensuring that each subcarrierexperiences a flat-fading channel. In conventional MIMOnetworks, only BB beamforming has been used to maximizethe SE, provided that each node has a fully connected RFchain corresponding to each antenna. However, in mmWavecommunications, the small aperture size of the antenna andthe large array size make it impossible for each antenna tohave an RF chain. Thus, hybrid precoding has been utilizedwith a much lesser number of RF chains than the numberof antennas, e.g., for gNB with 256 antennas, the number ofRF chains is set to 4. For the wideband channel, we assumethe BB beamforming is different for each subcarrier and isbased on the number of RF chains and that of data streams.In contrast, the RF beamforming is achieved via phase shifters(PSs) and is the same for all subcarriers. The dimension ofthe RF beamforming is defined by the number of RF chainsand the length of the antenna array. There are two kinds ofhybrid transceiver structures studied in [1], • Fully connected, where each RF chain connects to eachantenna, i.e., all the antennas are connected to each ofthe RF chains. • Partially connected (or subarray), where each RF chainonly connects to a disjoint subset of antennas.Although both structures employ fewer RF chains, the secondstructure is easier to deploy and more cost-efficient in practice.Since in a fully connected structure, mmWave antenna spacingand aperture size are small, which causes a high correlationbetween the outputs of RF chains. For the multiuser scenario, each subarray is set to serve a single user, which means thatthe number of subarrays can be selected based on the numberof users, see Fig. 3. In Fig. 3(b), each user is shown to beserved by 1 subarray with 16 antenna arrays panel.Fig. 3(a) gives the architecture of a multiuser hybridtransceiver for an IAB-FD wideband mmWave system, whichis used for the analysis for the IAB networks in this work. Forthe transmitter side, the OFDM block performs the inverseDFT (IDFT) and adds the CP to the precoded streams usingthe BB precoder. On the receiver side, the OFDM blockremoves the CP and performs the DFT, followed by the BBcombiner operation. Since each of the users (says U in total)communicates single data streams, the total number of datastreams ( N s = U ) should not exceed the number of RF chainsat the transmitter of the IAB-node.Our objective in hybrid transceiver design is to maximizethe SE across all subcarriers for access and backhaul links.This joint maximization problem concerning the RF and theBB precoders and combiners has a few constraints as follows.Since RF precoders and combiners are implemented usingPSs, it poses the constraint that the magnitude of each entryof the RF precoder and combiner matrices should be preciselyequal to 1. Further, the effective coupled RF and BB precodersmust satisfy the transmit power constraint. Assuming equalpower allocation across data streams, the squared norm ofthe hybrid precoder at each subcarrier should not exceedthe length of the data stream vector. Since the maximizationproblem is non-convex due to coupled RF and BB variables,a joint optimal solution for these variables is intractable.Interestingly, the near-optimal solution, where the RF andthe BB variables are obtained separately, is studied in [1].Ideally, the RF part of the hybrid precoders or the combinersis computed as the dominant eigenvector corresponding tothe eigenvalue decomposition (EVD) of the channel corre-lation matrix (i.e., the sample covariance matrix). In additionto this, the easier implementation of the subarray structuresimplifies the precoder and the combiner design to a blockdiagonal form, which incurs a lower computational com-plexity. Thus, for the subarray-based structure, RF variablesare obtained using the correlation matrix of the sub-channelmatrix corresponding to the antenna elements of the subarray.Note that EVD incurs a cubic computational overhead (say O ( N ) ). Thus, the subarray structure reduces the overheadto O (( N/U ) ) . However, the optimal solution above needs toaccess the channel state information (CSI) of the large MIMOchannel, which is unable to estimate in reality. Therefore,in this article, we assume the accurate knowledge of RFeffective channel only, where the RF precoders and combin-ers are provided by genie. In practice, these RF quantitiesare obtained by beam-training codebooks [10]. Next, theoptimal BB precoders/combiners can then be obtained asthe left/right dominant singular vectors of the estimated RFeffective channel matrix. Note that this BB transceiver designis applicable for the nodes, which have perfect interferencecancellation or do not experience interference. However, inthe IAB networks, there is strong SI presents at the IAB-node,needing cancellation. In thi case, the above hybrid design forthe IAB-node needs to be modified. ID,1 H SI IAB-node RXH GI gNBUN s ! ! ! ! ! !! ! RF PrecoderRFChainOFDM Block ! ! ! ! ! ! ! ! ! ! ! RFChainOFDM Block ! ! ! ! ! ! ! ! ! ! ! DigitalBasebandPrecoder ! ! ! ! ! ! RFChain OFDMBlock ! ! ! ! ! ! ! ! ! ! ! RFChain OFDMBlock ! ! ! ! ! ! ! ! ! ! ! DigitalBasebandCombiner ! ! ! ! ! !! ! ! ! ! !! ! RF Precoder
UEsIAB-node TX ! ! ! ! ! !! ! RF PrecoderRFChainOFDM Block ! ! ! ! ! ! ! ! ! ! ! RFChainOFDM Block ! ! ! ! ! ! ! ! ! ! ! DigitalBasebandPrecoder ! ! ! ! ! ! RFChain OFDMBlockRFChain OFDMBlock DigitalBasebandCombiner ! ! ! ! ! ! RF Precoder UN s N s H ID,U
DigitalBasebandCombiner ! ! ! ! ! ! ! ! ! ! ! User 1User U ! ! ! RF Precoder N s (a) IAB-node Tx User 1User 2User 3User 4 (b)Fig. 3. a) Multiuser hybrid transceiver for FD-IAB wideband mmWavesystem; b) subarray structure for multiuser transmission (8 × Multiuser Interference and Self-Interference Cancellation
To maximize the SE at the IAB-node receiver, the BB pre-coders/combiners at the IAB-node must achieve the following.The transceiver design should • mitigate the RSI at the receiver of the IAB-node, and • cancel the MUI at the transmitter of the IAB-node.Technically, in mmWave, such a high-power SI is likelyto exceed the limitation of the dynamic range on analog-to-digital converters (ADCs) and results in a strong non-linearsignal than that of desire signal. Therefore, the antenna andthe RF cancellation are adopted before the digital process tocancel out a large amount of SI [13]. However, the study in[3] states that for the mmWave wideband, the RF cancellationfaces difficulties in the canceler design due to the realizationof large number of taps and the high delay spread of the SIchannel, and also experiences severe performance degradationdue to RF impairments, as compared to that in the microwavecommunications. A wideband active analog SIC is studiedin [14]. With this novel RF cancellation technique, thosedifficulties on traditional RF canceler design can be overcome.Consequently, the remaining RSI will be handled by thedigital cancellation, i.e., by applying the MMSE BB combinerat the IAB-node. In order to achieve a good digital SIC, thenumber of RF chains at the receiver of the IAB-node shouldbe at least the sum of the number of data streams transmittedand received by the IAB-node. Since the BB SIC depends on the estimated CSI of the SI channel, the CEE has a strongimpact on the performance of digital SIC. A staged SIC whichcombines the RF and the digital cancellation is studied in ourrecent work [10]. Regarding the MUI, traditional ZF is utilizedat the IAB-node transmitter based on the RF effective channelto obtain the desired BB precoder. RF Insertion Loss
The RFIL, L RF , which is caused by PSs, power dividers(PDs), and power combiners (PCs), is an important loss thatcannot be easily compensated by the existing technologiesin the mmWave. Failure to take account of the RFIL mayresult in higher analytical spectral efficiency. To act the impactof the RFIL, the factor, / √ L RF , is multiplied with the RFprecoder/combiner matrices.For the fully connected structure, the RF precoding requires N RF PDs ( N t -way), N t PCs ( N RF -way) and N t N RF PSs,while the RF combining needs N r PDs ( N RF -way), N RF PCs ( N r -way), and N r N RF PSs, where N t , N r and N RF denotes the number of transmitters, receivers, and RF chains,respectively.On the other hand, for the RF precoding with U subarrays, U PDs ( N t /U -way) and N t PSs are required, while at eachsubarray (user) of the receiver, U PCs ( N r /U -way) and N r PSs are needed. Specially, at the receiver of the IAB-node, N r PDs ( N RF /U -way), N RF N r /U PSs, and N RF PCs ( N r /U -way) are required.Given that a cascade of (cid:100) log ( X ) (cid:101) stages of 2-way PDs and (cid:100) log ( Y ) (cid:101) stages of 2-way PCs are utilized to construct the X -way PD and the Y -way PC, respectively. L RF is given bythe product of the static power loss of PDs (i.e., P D (cid:100) log ( X ) (cid:101) dB), PSs (i.e., P PS dB), and PCs (i.e., P C (cid:100) log ( Y ) (cid:101) dB),where P D = 0 . dB and P C = 3 . dB denote the powerloss of the PD and the PC, respectively. Moreover, there aretwo kinds of PSs, i.e., the active PS ( P PS = − . dB) andthe passive PS ( P PS = 8 . dB) [15].S IMULATION R ESULTS
In this section, simulations are presented to analyze the SEfor our hybrid precoding design with the impact of the CEEand the RFIL. The OFDM system has K = 512 subcarriers,where each channel realization has D = 128 delay taps. Fora 4-subarray (user) hybrid precoding system, each subarray(user) has × UPA with 1 RF chain and 1 data stream.
10 -5 0 5 10 15
SNR (dB) S p ec t r a l E ff i c i e n c y ( b it s / s / H z ) FD-Fully connected (w/o SI)HD-Fully connectedFD-Subarray (w/o SI)HD-Subarrayw/o RFIL (w/ SIC)w/ RFIL active PSs (w/ SIC)w/ RFIL passive PSs (w/ SIC) (a) -10 -5 0 5 10 15
SNR (dB) S p ec t r a l E ff i c i e n c y ( b it s / s / H z ) FD-Fully connected (w/o SI)HD-Fully connectedFD-Subarray (w/o SI)HD-Subarrayw/o RFIL (w/ SIC)w/ RFIL active PSs (w/ SIC)w/ RFIL passive PSs (w/ SIC) (b)Fig. 4. Comparison of the impact of RFIL on the SE of a 4-user mmWave-IAB-FD system with different hybrid precoding structures in terms ofdifferent kinds of PSs. The number of subarrays is equal to that of the user: a)backhaul link: 16 ×
16 UPA, 4 (8) RF chains at Tx (Rx), 4 data streams; b)access link: 16 ×
16 UPA and 4 RF chains at the Tx. Each user is equippedwith 1 RF chain and 4 ×
16 UPA and receives 1 data stream from Tx.
For successful digital cancellation, each subarray has 2 RFchains at the receiver of the IAB-node. We assume that an80 dB SIC has been applied before the digital cancellation bythe antenna and the analog cancellation [10]. We define SNR (cid:44) P r /σ , where P r = P t / ¯ P L is the ratio between transmitpower and average path loss according to the Friis’ law, and σ denotes the Gaussian noise power. A. Effect of RF Insertion Loss
Fig. 4 shows the SE of both the backhaul and the accesslink with different hybrid precoding schemes by comparingFD and HD transmission in the presence of the RFIL in termsof different kinds of PSs. Both subfigures show a similar trend.Without considering the impact of the RFIL, the SE with FD transmission of the fully connected structure is much higherthan that of the subarray structure, which has the differenceof around 20 bits/s/Hz and 12 bits/s/Hz for the backhaul andthe access links, respectively, at SNR = 15 dB. For the HDscheme, this difference reduces to a half. However, in thepresence of the RFIL, the SE obtained from the subarraystructure is close to that given by the fully connected one,which means that our precoding scheme experiences lesseffect from the RFIL. Moreover, it can be seen that the use ofactive PSs can provide a higher SE than that with passive PSs;however, with more power consumption [15]. Specifically, forthe backhaul link with ideal RF components, the SE of FDwith SIC is close to the ideal one (i.e., with perfect SIC),which indicates the successful SIC.
B. Effect of Channel Estimation Error
We assume that only the RF effective SI channel is knownwith uncertainty. Therefore, only the backhaul link perfor-mance will be affected by the CEE. From Fig. 5, it canbe observed that irrespective of the selection of PSs, thehigher SNR shifts the SE intersection of FD and HD tothe left. At the right of the intersection, the FD schemehas less SE as the HD due to the higher CEE. Moreover,compared with the fully connected structure, our subarray-based hybrid precoding scheme is more sensitive to the CEE.Therefore, more advanced techniques are needed to estimatethe RF effective SI channel as accurately as possible. Further,interestingly, with passive PSs, the intersection points shifts tothe right, as compared with that for active PSs, implying themore tolerance of the system with passive PSs. It can be notedthat although the fully connected structure shows a better SE,yet the incurred hardware cost is much less for the subarraystructure.
C. Effect of RF Chains on Digital SIC
In Fig. 6, the digital SIC ability in terms of the SE ofthe backhaul link is plotted with different numbers of RFchains at the IAB-node receiver. The fully connected hybridprecoding schemes are assumed to have 4 (8) RF chainsat the transmitter (receiver). The ideal curves are plottedby assuming perfect SIC. It is evident that the ideal fullyconnected-based precoding provides a close performance tothe ideal full digital scheme, and leaves a gap with respectto the ideal subarray-based precoding scheme. Regardingthe digital cancellation ability of the subarray structure, themore RF chains at the receiver of the IAB-node, the moreimprovements in the SE can be seen and the smaller theSE difference with respect to the ideal subarray curves. At15 dB SNR, with different numbers of RF chains at thereceiver of the IAB-node ( L = 2 , , per subarray), the SE ofdeploying digital SIC is improved nearly 23%, 33%, and 34%,respectively, and the corresponding rate loss gets to around4.7, 2.1, and 1 bit(s)/s/Hz, respectively.C ONCLUSION
In this article, we have presented the multiuser mmWave-FD-IAB architecture according to the latest 3GPP standard
60 -50 -40 -30 -20 -10 0
Estimation Error e2 (dB) S p ec t r a l E ff i c i e n c y ( b it s / s / H z ) FD-Fully connected w/ SICHD-Fully connectedFD-Subarray w/ SICHD-Subarrayw/ RFIL active PSsw/ RFIL passive PSs
SNR = -10 dB, 0 dB, 10 dB (a) -60 -50 -40 -30 -20 -10 0
Estimation Error e2 (dB) S p ec t r a l E ff i c i e n c y ( b it s / s / H z ) FD-Fully connected w/ SICHD-Fully connectedFD-Subarray w/ SICHD-Subarrayw/ RFIL active PSsw/ RFIL passive PSs
SNR = -10 dB, 0 dB, 10 dB (b)Fig. 5. Comparison of the impact of CEE on the backhaul link SE in the presence of RFIL of a 4-user mmWave-IAB-FD system with different hybridprecoding structures in terms of different SNR values. Equipped with 16 ×
16 UPA, 4 (8) RF chains at Tx (Rx), 4 data streams are transmitted. The numberof subarrays is equal to that of the user: a) with active PSs; b) with passive PSs. -10 -5 0 5 10 15
SNR (dB) S p ec t r a l E ff i c i e n c y ( b it s / s / H z ) FDHDIdeal full digitalIdeal fully connectedIdeal subarray (L=8)Ideal subarray (L=4)Ideal subarray (L=2)Subarray w/ digital SICSubarray w/o digital SIC
Fig. 6. Digital SIC ability of a 4-user wideband mmWave-FD-IAB subarraynetwork in terms of different numbers of RF chains ( L = 2 , , on eachRx subarray. Equipped with 16 ×
16 UPA on both sides, 4 RF chains at Txand 4 data streams are transmitted. for the IAB networks. Wideband and FD operations havebeen investigated towards the SE perspective. Further, thegeneral mmWave channel model is described, followed bythe characterization of the SI channel for mmWave FD op-eration, including the challenges in the SI channel estima-tion. Through a hardware cost-effective and computationallyefficient subarray-based hybrid precoding scheme, with theobjective of SE maximization in the IAB networks, MUI andRSI are mitigated at the IAB-node transmitter and receiverusing the BB ZF and MMSE, respectively. The impact ofthe RFIL with active or passive PSs has been analyzed. To observe the effect of the imperfect RF effective CSI, the SEis plotted for different values of CEE in the presence of theRFIL, and compared with the HD operation. Simulations haveshown that if the CEE is inversely proportional to SNR, theimprovement of FD and HD can be observed. Moreover, thesystem with passive PSs can tolerate higher CEE than thesystem with active PSs.Since the subarray hybrid precoding scheme is sensitiveto the CEE, adjustments need to be investigated for accurateRF effective SI channel estimation. Further, equal powerallocation assumption can be relaxed, and optimal power canbe allocated to the effective channel. In practice, the PSsare not continuously controlled. Therefore, we will focus onquantization schemes with an efficient codebook design inthe future. Moreover, an efficient antenna and RF cancellationare important to investigate to leverage the advantages of FDtransmission. R
EFERENCES[1] S. Park, A. Alkhateeb, and R. W. Heath, “Dynamic Subarrays forHybrid Precoding in Wideband mmWave MIMO Systems,”
IEEE Trans.Wireless Commun. , vol. 16, no. 5, pp. 2907–2920, May 2017.[2] O. Teyeb et al. , “Integrated Access Backhauled Networks,” in
Proc.IEEE 90th Vehicular Technology Conference , Honolulu, HI, USA, Sep.2019.[3] Z. Xiao, P. Xia, and X. Xia, “Full-Duplex Millimeter-Wave Commu-nication,”
IEEE Wireless Commun. Mag. , vol. 24, no. 6, pp. 136–143,Dec. 2017.[4] B. Lee et al. , “Reflected Self-Interference Channel Measurement formmWave Beamformed Full-Duplex System,” in
Proc. IEEE GlobecomWorkshops , San Diego, CA, USA, Dec. 2015.[5] 3GPP, “NR; Study on Integrated Access and Backhaul,”
TR 38.874 (Rel.16) , Dec. 2018.[6] ——, “NG-RAN; Architecture Description,”
TS 38.401 (Rel. 16) , Sep.2020.[7] O. Taghizadeh et al. , “Environment-Aware Minimum-Cost WirelessBackhaul Network Planning with Full-Duplex Links,”
IEEE Syst. J. ,vol. 13, no. 3, pp. 2582–2593, Sep. 2019.8] M. Polese et al. , “Integrated Access and Backhaul in 5G mmWaveNetworks: Potential and Challenges,”
IEEE Commun. Mag. , vol. 58,no. 3, pp. 62–68, March 2020.[9] A. Alkhateeb and R. W. Heath, “Frequency Selective Hybrid Precodingfor Limited Feedback Millimeter Wave Systems,”
IEEE Trans. Com-mun. , vol. 64, no. 5, pp. 1801–1818, May 2016.[10] J. Zhang et al. , “Design and Analysis of WidebandFull-Duplex FR2-IAB Networks,”
IEEE Trans. Commun. ,under review. [Online]. Available: https://uoe-my.sharepoint.com/:f:/g/personal/s1804540 ed ac uk/EuiTu5H9wnZP \ pjEMB5eARuYBDOhJg2DlzNgxS3L5wCMhMg?e=v1B4c1[11] K. Abbas et al. , “Joint Channel and Phase Noise Estimation formmWave Full-Duplex Communication Systems,” EURASIP J. Adv.Signal Process. , vol. 18, pp. 1–12, March 2019.[12] C. Masouros, M. Sellathurai, and T. Ratnarajah, “Vector PerturbationBased on Symbol Scaling for Limited Feedback MISO Downlinks,”
IEEE Trans. Signal Process. , vol. 62, no. 3, pp. 562–571, Feb. 2014.[13] L. Song, Y. Li, and Z. Han, “Resource Allocation in Full-Duplex Com-munications for Future Wireless Networks,”
IEEE Wireless Commun.Mag. , vol. 22, no. 4, pp. 88–96, Aug. 2015.[14] H. Luo, M. Holm, and T. Ratnarajah, “Wide-Band Active Analog Self-Interference Cancellation for 5G and Beyond Full-Duplex Systems,” in
Proc. 54th Asilomar Conference on Signals, Systems and Computers ,Pacific Grove, CA, USA, Nov. 2020.[15] L. N. Ribeiro et al. , “Energy Efficiency of mmWave Massive MIMOPrecoding with Low-Resolution DACs,”
IEEE J. Sel. Topics SignalProcess. , vol. 12, no. 2, pp. 298–312, May 2018. A CKNOWLEDGMENT
The work was supported in part by the research grant fromHuawei Technologies (Sweden) AB.B
IOGRAPHIES J UNKAI Z HANG [S’21] received his B.Eng. degree inCommunication Engineering from Shenyang LigongUniversity, Shenyang, China, in 2018, and M.Sc. degree inSignal Processing and Communications, with Distinction,from The University of Edinburgh, Edinburgh, UK., in 2019.He is currently with the Institute for Digital Communications,The University of Edinburgh, Edinburgh, UK., as a Ph.D.candidate. His research interests include 5G and beyondwireless networks, millimeter-wave communications, full-duplex radio, and Massive MIMO.N
AVNEET G ARG [S’15, M’19] received the B.Tech. degreein electronics and communication engineering from Collegeof Science & Engineering, Jhansi, India, in 2010, and theM.Tech. degree in digital communications from ABV-IndianInstitute of Information Technology and Management,Gwalior, in 2012. He completed the Ph.D. degree in June2018 from the department of electrical engineering at theIndian Institute of Technology Kanpur, India. From July2018-Jan 2019, he visited The University of Edinburgh,UK. From February 2019-2020, he was employed as aresearch associate in Heriot-Watt university, Edinburgh, UK.Presently, he is working as a research associate in TheUniversity of Edinburgh, UK. His main research interestsinclude interference alignment, edge caching, optimization,and machine learning.M
ARK H OLM [S’98, M’01] received his B.Sc. (hons) inLaser Physics and Optoelectronics from the University ofStrathclyde in 1997, before studying for his Ph.D. in Physicsfrom the University of Strathclyde which he received in 2001. He currently works as a technical lead and hardwaresystem architect for Huawei Technologies (Sweden) AB withinterests in microwave radio, phased array antennas, fullduplex radio systems, and photonic radios. In the past, hewas the microwave lead on AESA Radar systems, SeniorEngineer responsible for GaAs pHemt modeling, and alsolaser and package design engineer for SFP/XENPACK Fibremodules. He is published in the fields of laser design, andGaAs device modeling.T
HARMALINGAM R ATNARAJAH [S’94, A’96, M’05, SM’05]is currently with the Institute for Digital Communications,The University of Edinburgh, Edinburgh, U.K., as a Professorin digital communications and signal processing. He hassupervised 15 Ph.D. students and 21 postdoctoral researchfellows and raised more than 11+ million USD of researchfunding. He was the Coordinator of the EU projects ADEL(3.7M (cid:164) ) in the area of licensed shared access for 5G wirelessnetworks, HARP (4.6M (cid:164) ) in the area of highly distributedMIMO, the EU Future and Emerging Technologies projectsHIATUS (3.6M (cid:164) ) in the area of interference alignment, andCROWN (3.4M (cid:164)(cid:164)