Contention-based Grant-free Transmission with Independent Multi-pilot Scheme
CContention-based Grant-free Transmission withIndependent Multi-pilot Scheme
Zhifeng Yuan, Weimin Li, Zhigang Li, Yuzhou Hu, Yihua MaZTE Corporation, Shenzhen, ChinaEmail: {yuan.zhifeng, li.weimin6, li.zhigang4, hu.yuzhou, yihua.ma}@zte.com.cn
Abstract —Contention-based grant-free transmission is verypromising for future massive machine-type communication(mMTC). In contention-based transmission, the random pilotcollision is a challenging problem. To solve this problem,multiple pilots scheme is used to reduce the pilot collisionprobabliltiy. However, the existing work on multiple pilotsrelies on the low correlation of spatial channels, limiting itsapplicability. In this paper, an independent multi-pilot schemeis proposed, which utilizes the diversity of multiple pilots and isnot limited by the spatial correlation. The receiver employsinterference cancellation for both data symbols and multiplepilots to ensure the performance. The simulation results alsoshow that the proposed independent multi-pilot scheme cansignificantly improve the BLER performance and increase thenumber of simultaneous access users.
Keywords — grant-free, independent multi-pilot, pilot collision,contention-based transmission I. I
NTRODUCTION
In beyond 5G (B5G) wireless communication system,massive Machine-Type Communication (mMTC) will beone of the most important scenarios [1, 2]. The mMTCscenario features massive access devices, small data packets,low transmission rate and sporadic communication. Inaddition, the access devices should be characterized by lowcomplexity, power-saving and low cost. The grant-freeaccess scheme [3, 4], which allows user equipment (UE) totransmit data autonomously without the need to sendscheduling request and wait for dynamic scheduling, hasbeen proved to be a promising solution to satisfy therequirements of mMTC scenario with the followingadvantages: (1) saving signaling overhead; (2) reducingtransmission latency: The complex random access andresource grant procedure are no longer needed; (3) reducingpower consumption: The terminal devices can be in idle statefor a long time, and can immediately switch to transmittingstate when data arrives. At low transmission rate, non-orthogonal multiple access (NOMA), with the ability toexploit near-far advantage, has a higher capacity thanorthogonal multiple access, and can support more accessdevices on the same transmission resources [5, 6]. Thecombination of grant-free and NOMA can solve theproblems of connection density, signaling overhead, terminalcomplexity and power consumption, which is suitable to theB5G mMTC scenario.In the mMTC scenario with massive connections andsporadic small data packets, reserving dedicated transmissionresource for each connection is unrealistic, resulting indifferent UEs sharing the same wireless resource block. Inautonomous grant-free transmission, each UE autonomouslyselects transmission resources including pilot sequence in acontention-based manner. Inevitably, multiple UEs mayselect the same pilot sequence known as pilot collision, which will severely degrade the system performance [7, 8].For traditional transmission scheme where a single pilotsequence and a data payload are included, the pilot collisionprobability can be expressed as [9]
KKN
NAP (1)where N is the size of the defined pilot pool containingorthogonal pilot sequences, K is the number of concurrentUEs, and !/ ! KN A N N K is the number of permutationsof N items taken K at a time. It can be derived that, as thenumber of concurrent UEs K increases, the pilot collisionprobability rises rapidly.Increasing the number of pilot sequences to reduce thepilot collision probability is a common approach [10, 11].However, as the number of orthogonal pilot sequencesincreases, the length of the pilot sequence also needs to beincreased, which consumes more time-frequency resources.With limited time-frequency resources, the resourcesavailable for data are less, affecting the efficiency of datatransmission. In addition, the complexity of blind detectionfor a longer pilot sequence during active user detection(AUD) also increases significantly. Another approach tosolve the pilot collision problem is a data-only based grant-free transmission scheme [12, 13], which directly removesthe pilot signal, resulting in no pilot collision and overhead.In this scheme, advanced blind multi-user detection (MUD)techniques without pilot-based channel estimation, such asblind receive beamforming without spatial channelinformation, blind activity detection based on the secondorder moment of the data symbols, bind MMSE de-spreading,and blind equalization via partition-matching method, arerequired.In this paper, an independent multi-pilot (IMP) scheme isproposed, which can significantly reduce the pilot collisionprobability, with moderately increased blind detection efforts.Different from the traditional single pilot (TSP) scheme, theproposed scheme uses multiple pilot sequences which areselected or generated independently, thus uncorrelated, andcollision occurring on all pilots is obviously lower than thatin the TSP scheme under the same pilot resource overhead.Unlike the multiple pilots scheme relying heavily on the lowcorrelation of the accessing users' spatial channels [14],which limits its applicability, the proposed IMP schemeworks for a more general case without any requirement ofspatial correlation. The simulation results show that both theBlock Error Rate (BLER) performance and the number ofaccess UEs are significantly improved at the cost ofincreased decoding attempts per user. The article isorganized as follows. In Section II, the independent multi-pilot scheme including the multi-pilot based transmitter andreceiver is elaborated in detail, then the pilot collisionprobability, channel estimation precision, receiveromplexity, and the number of independent pilots are alsodiscussed. Section III provides the performance evaluationresults together with the comparison between theindependent multi-pilot scheme and traditional single pilotscheme. The conclusion is presented in Section IV.II. T HE I NDEPENDENT M ULTI -P ILOT S CHEME
A. Multi-pilot and Transmitter Design
Different from the traditional scheme configured withsingle pilot, the IMP scheme uses multiple independentpilots under the same pilot resource overhead. One typicalresource sharing by the multiple pilots is shown in Fig. 1,where each pilot occupies a disjoint sub resource block of thepilot resource. In principle, the lengths or the sequence typesof the pilots are not necessarily identical. However, multiplepilots of equal length and identical sequence type arebeneficial to minimize the collision probability, channelestimation error and implementation complexity.
Fig. 1. a) Traditional single pilot scheme; b) Independent multi-pilot scheme
Taking w independent pilots of equal length and identicalsequence type as an example, each UE independently selects w pilot sequences from a predetermined pilot pool and mapsthem to the transmission resources together with the datasymbols in one-shot transmission. The total overhead of w independent pilots is the same as that of the single pilot intraditional scheme. Assuming that the size of the pilot poolof the TSP scheme is N TSP (i.e., N TSP orthogonal pilotsequences of length N TSP ), and the w independent pilots usedisjoint resources, the size of the pilot pool of the IMPscheme N IMP is therefore N TSP / w (i.e., N TSP / w orthogonal pilotsequences of length N TSP / w ).For IMP-based transmission, the probability of collisionon all w pilots P IMP would be significantly lower than that inTSP-based scheme P TSP provided N TSP is not too small,which is a reasonable assumption because N TSP in generalneeds to be large enough to ensure an acceptable collisionrate for traditional scheme. Considering the case of two UEsrandomly selecting pilot sequences from the predefined pilotpool, P TSP is 1/ N TSP , P IMP is (1/ N IMP ) w = ( w / N TSP ) w . Assumingthat two independent pilots are utilized, i.e. w = 2, P IMP / P TSP is 4/ N TSP , which means that IMP-based scheme has lowercollision probability when N TSP is larger than 4, and theprobability would be rapidly reduced with the increasing of N TSP . Therefore, for IMP-based transmission, the detection ofa given UE's data can depend on at least one of its pilots thatdoes not collide with other UEs', and better performance canbe expected due to the lower pilot collision probability.To improve the detection performance of other UEs whosepilots are collided with the UEs that have been decoded,interference cancellation (IC) on pilot signals is required tobe implemented. That means the receiver should know thepilot sequences the UE used. To realize this, putting theinformation of multiple pilot sequences into the data payloadseems a straightforward solution, such that once a given UE'sdata is successfully decoded, all the pilot sequences used bythe UE can be determined and then IC on the multiple pilots can be performed. However, this solution would incuroverhead. Alternatively, some given bits in codeword, whichin general are independent or uncorrelated of each other, canbe used to determine the multiple pilot sequences. Concretely,if N IMP =2 m , w * m coded bits can be used to select the w independent pilot sequences from the pilot pool. In this way,once the coded bits pass the CRC check, the indexes of w pilot sequences can be determined. B. Receiver Design
The receiver of the IMP-based grant-free transmissionscheme is mainly composed of two parts: blind MUD andcodeword level IC. The receiver has no prior information onthe number of access UEs and the pilots selected by theseUEs. Therefore, the receiver has to perform blind detectionincluding AUD and channel estimation, MMSE equalization,demodulation and decoding, etc., as shown in Fig. 2.
Fig. 2. Receiver processing diagram
For convenience, the subsequent description takes twoindependent pilots as an example. The receiver can performblind MUD by using two independent pilots in parallel, andthen perform IC on both the pilots and data symbols, asshown in Fig. 3a). It also can use one pilot to perform MUDand IC at first, then use another pilot to perform MUD andIC, i.e. two pilots are processed serially, as shown in Fig. 3b).If parallel processing procedure shown in Fig. 3a) is adopted,one UE may be successfully decoded based on both pilots.However this UE's signal can not be cancelled twice, thereceiver therefore has to determine which channel estimationderived from P1 or P2 should be used to reconstruct the UE’sreceived symbols. The receiver can also process the channelestimations on both pilots to obtain a weighted channelestimation for reconstruction. In another way, transmittedsymbols of all decoded UEs can play the role of ‘pilot’ toderive refined channel estimations, which then can be usedfor more accurate received symbols reconstruction. Thismethod can be called data-aided channel estimation, which isquite beneficial for contention-based grant-free transmissionand will be discussed in detail in latter sections.
Fig. 3. a) Parallel detection based on two pilots; b) Serial detection based ontwo pilots
Assume K active UEs are simultaneously transmitting datain contention-based grant-free manner, and each UEindependently selects two pilot sequences from the definedorthogonal pilot resource pool of size N IMP , IMP , ,..., , N Z z z z
IMPIMP . NN C z The received pilotsymbols on the i- th pilot can be expressed as , , ,1 , 1, 2 Ki i k n i k ik h i y z n (2)where , , n i k z represents that UE k randomly selects a pilotsequence n z on the i- th pilot, n Z z ; , i k h is the channelcoefficient of UE k on the i- th pilot, here flat fading channelis assumed for simplicity; i n is the additive white Gaussiannoise (AWGN), ~ (0, ) i CN I n .1) Blind detectionBy a hypothesis testing on all possible pilot sequences,blind detection of active users' pilot sequences can beperformed, which can be expressed as , , ,1 , 1,..., Kx i i k x n i k x i IMPk h x N z y z z z n (3)where ( ) represents conjugate transpose. By using theorthogonality of pilot sequences, i.e., x n x n z z , thefollowing can be derived for UE k , n i i k n n n i h z y z z z n (4)Then normalized detection result, which is also the channelestimation of UE k on the i- th pilot, can be obtained by , , ˆ n i n ii k i kn n n n h h z y z nz z z z (5)By setting appropriate detection threshold, active user on the i- th pilot can be identified based on the normalized detectionresults. The threshold should be set considering the influenceof interference and noise, and to realize a trade-off betweenfalse alarm and miss detection. Assuming that S users areidentified, based on the channel estimations of these users,MMSE based equalization can be implemented to derive thedetected data symbols of each user, which then can be sent todemodulator and decoder to obtain the data bits transmittedby corresponding users. CRC check can be used here tojudge whether the data of a user is correctly decoded. Asdiscussed before, information of multiple pilot sequences aswell as identification number of a user can be carried in thedata payload. So from the correctly decoded data bits, accessuser can be determined together with its multiple pilots,which would help the interference cancellation process.2) Interference CancellationPilot collision has a strong impact on the performance ofcontention-based grant-free transmission. If two or more UEsselect the same pilot sequence, only one pilot sequence canbe detected, and the channel estimation based on this pilotwould be the sum of channels experienced by these UEs. Ifthere are large disparity between the collided UEs' receivedpower, the strongest UE may be successfully decoded. Inorder to detect the remaining UEs, the data symbols as wellas the pilot signal of the successfully decoded UEs should bereconstructed, and cancelled from the received symbols, thenblind MUD can be performed iteratively.For IMP-based scheme, it should be noted that IC shouldbe performed for all pilots. The two pilot sequences selected by a decoded UE can be determined according to the 2 m bitsin the data part . Assuming that Q UEs are successfullydecoded, and taking serial detection based on two pilotsshown in Fig. 3b) as an example, the
IC for pilots and datacan be respectively expressed as
Q , , ,1 - , 1, 2 p p i q v p qq h p y y z (6) Qd d ,1 - i q qq h y y d (7)where , , v p q z represents the pilot sequence v z used by UE q on p- th pilot, v Z z ; , i q h is the filtered channel estimationbased on , ˆ i q h of UE q on the i- th pilot, it means that channelestimation on the i- th pilot would be used for cancellation onboth pilots as well as data part; q d is the reconstructedmodulated data symbols of UE q .Based on p y and d y , the next round of blind MUD can beperformed. The process can be iterated until no active UEcan be identified or no new UE can be successfully decoded.Generally, the accuracy of pilot-based channel estimationis limited by the pilot power, which is more obvious forIMP-based scheme because each pilot’s power is reduced. Itwould cause large residual error and affect the detection ofremaining UEs. In addition, as mentioned above, when pilotcollision exists, the channel estimation is the sum of channelsof muiltiple UEs, IC based on that would be quite inaccurate,and the remaining weaker collided UEs may not be detectedsubsequently, resulting in miss detection.To improve the accuracy of IC, data-aided channelestimation can be utilized, where all the transmitted symbols(including the pilot symbols and modulated data symbols) ofdecoded UEs can be employed as ‘pilot’ to derive refinedchannel estimations for these UEs through least squaremethod [12]. It can be expressed as -1 ˆ h = (D D) D y (8) where y is a vector composed of the received symbols, D = [D , D , ... , D ] is a matrix composed of the reconstruct-ed transmitted symbols of all the Q decoded UEs till thisround, ˆ ˆ ˆ ˆ, , , T h = [h h ... h ] is a vector composed of thechannel estimations for these Q UEs.It can be seen from [13] that data-aided channel estimationwould be more accurate with the increasing number ofsuccessfully decoded UEs, and it is more suitable for the casewith pilot collision as the data of different UEs can beconsidered uncorrelated. Therefore, data-aided channelestimation in IC can help to minimize the residual error andreduce the miss detection rate, which are particularlyimportant for the weaker UEs.
C. Pilot Collision Probability
More discussions are provided here to illustrate theadvantages of IMP scheme in reducing pilot collisionprobability. The case with two UEs has been discussedbefore, while for the case with more UEs, the pilot collisionanalysis would be a little complex. Fig. 4 shows a case with UEs ( K = 3), wherein UE2 collides with others on bothpilots, such that it cannot be decoded in the first round ofblind MUD. However, UE3 could be successfully decodedbased on its first pilot ‘z4’ which does not collide with others.Then the second pilot ‘z5’ of UE3, although collides withthat of UE2, can be determined according to the 2 m bits inthe decoded data. Similarly, UE1 could be successfullydecoded based on its second pilot ‘z7’. After cancellation forUE1 and UE3, the receiver can perform next round ofdetection, where UE2 would be decoded. It can be seen that,although UE2 collides with others on both pilots, it finallycan still be decoded. This case can be regarded as a solvablepilot collision case, and will not affect the BLERperformance considerably. Fig. 4. One case of pilot collision with 3 access UEs
Two UEs colliding with each other on both pilots is themain case that affects the performance. For example, if UE1shown in Fig. 4 selected ‘z5’ on pilot P2, only UE3 could besuccessfully decoded, but both UE1 and UE2 can not bedecoded as they collide with each other on both pilots. Thecollision probability of this case for K = 3 can be expressedas ( 2 ) / KK N N K N N
P C A A C A A N (9)where N is the size of the pilot pool of the IMP scheme, i.e., N = N IMP , and !! ! KN NC K N K is the number of combinationsof N items taken K at a time. Further, Eq. (9) can beapproximated as K P C N (10) The probability in Eq. (10) can be regarded as the number ofcombinations for randomly selecting two UEs from threeconcurrent UEs multiplied by the probability of collision onboth pilots for the case with two UEs. Generalizing Eq. (10)to K > 3 UEs seems reasonable.Assuming that the size of pilot pool of the TSP scheme is N TSP = 24, and the size of pilot pool of the IMP schemeconfigured with two pilots is N IMP = 12, the pilot collisionprobability is about 12% for TSP scheme with 3 concurrentUEs according to Eq. (1), while according to Eq. (10) thepilot collision probability would reduce to 2%. Theprobability of three UEs colliding with each other on bothpilots at the same time is negligible. For more UEs, the pilotcollision probability can be analyzed similarly.
D. Accuracy of Channel Estimation
For IMP scheme, the total overhead and energy of w pilotsis the same as that of single pilot in the TSP scheme.Therefore, the energy of one pilot in the IMP scheme is only1/ w of that in the TSP scheme, which leads to lower channelestimation accuracy. Specifically, for the case with two pilots,the channel estimation accuracy will decrease by 3dB; and for the case with three pilots, the channel estimationaccuracy will decrease by 4.77dB. To balance the channelestimation accuracy and demodulation performance, transmitpower boost could be considered on the multiple pilots forIMP-based transmission. E. Receiver Complexity
As the receiver needs to perform MUD based on each pilotfor IMP scheme, the complexity would be increased. If theparallel detection procedure shown in Fig. 3a) is employed,the receiver may repeatedly decode the same UE on multiplepilots. While for the serial detection procedure shown in Fig.3b), by using interference cancellation after detection oneach pilot, decoding the same UE on multiple pilots can beavoided in some extent, which would save complexity. Somecomplexity comparison results are provided in Section III forreference.
F. Number of Independent Pilots
The number of independent pilots is closely related to thepilot collision probability, channel estimation accuracy andreceiver complexity. As discussed above, increasing thenumber of pilots can reduce the pilot collision probabilityand improve the system performance in collision-limitedscenarios. However, the more the number of pilots, the lowerthe accuracy of channel estimation. This will not onlydegrade the demodulation performance, but also lead to largeresidual error of interference cancellation, which directlyaffects the detection of the weaker UEs. The receivercomplexity also increases with the number of pilots.Therefore, the determination of the number of pilots requiresa comprehensive consideration of the system performanceand complexity. III. N UMERICAL R ESULTS
The performance of the proposed IMP scheme is evaluatedby link-level simulation, and the comparison with the TSPscheme is also provided. The simulation parameters areshown in Table 1.
TABLE I. S
IMULATION A SSUMPTION
Parameter ValueCarrier frequency 0.7 GHzSystem bandwidth 10 MHzAllocated bandwidth 6 PRBsModulation and Coding scheme QPSK, LDPCSize of transport block 160 bitsLength of CRC 16 bitsAntenna configuration 1Tx, 2RxChannel model and delay spread TDL-A 30nsReceiver algorithm MMSE-IC
The allocated transmission resource is 6 consecutivephysical resource blocks (PRBs) in the frequency domain,and 14 OFDM symbols in the time domain. There are 12subcarriers per PRB and subcarrier spacing is 15 kHz. TheDemodulation Reference Signal (DMRS) structure of TSPscheme can refer to DMRS type 2 in 5G new radio (NR).The difference is that the DMRS structure of TSP schneme ismore sparser in the frequency domain for a larger number oforthogoal reference signals to increase the multiplexingcapacity or support more access UEs. Multipe pilots of IMPscheme are mapped to independent, non-overlapping, andequal-sized sub-resource blocks in the frequency domain.Two kinds of pilot overhead are considered. In the first one, OFDM symbols are for pilot, i.e. the pilot overhead is 1/7,thus 24 orthogonal reference signals can be configured oneach PRB for TSP scheme, while for IMP scheme, 2 pilotswith 12 orthogonal reference signals per pilot can beconfigured on each PRB, or 3 pilots with 8 orthogonalreference signals per pilot can be configured on each PRB. Inthe other one, 4 OFDM symbols are for pilot, i.e. the pilotoverhead is 2/7, and the number of orthogonal referencesignals doubles for each scheme. The channel model is TDL-A with delay spread of 30 ns. The long term receiving SNRof all access UEs are equal. In the receiver, the serialdetection procedure shown in Fig. 3b) is used, and pilotbased realistic channel estimation is employed in thedetection and interference cancellation. To reduce the ICresidual error and improve the detection performance,reconstructed data aided channel estimation is alsoconsidered for IC in the simulation.
A. Pilot-based Channel Estimation Used in IC
In this subsection, pilot-based channel estimation is usedin IC. The comparison of the BLER performance betweenthe TSP scheme and the IMP scheme configured with 2pilots under 1/7 pilot overhead is shown in Fig. 5. From theresults, it can be seen that, the BLER performance of theIMP scheme is better than that of the TSP scheme in general.At BLER=0.1, the IMP scheme can support more than 8simultaneous access UEs, which is about twice the numberof UEs supported by the TSP scheme. For 4 UEs, the IMPscheme has a performance gain of more than 3 dB comparedwith the TSP scheme at BLER=0.1. For 6 to 10 UEs and atSNR less than -2 dB, the performance of the TSP schemewill be slightly better than the IMP scheme due to channelestimation accuracy. With the increasing of SNR, thereobviously exists error floor for the TSP scheme. The mainreason is the pilot collision probability increases as thenumber of UEs increases, which seriously affects theperformance of the TSP scheme. Thanks to the lower pilotcollision probability, the IMP scheme can significantlyimprove BLER performance and increase number of accessUEs.
Fig. 5. Comparison of BLER performance between the TSP scheme andthe IMP scheme with 2 pilots under 1/7 pilot overhead.
More comparisons of performance between the TSPscheme and the IMP scheme configured with 2 pilots areprovided in Fig. 6, where BLER = 0.1 and 0.01 are focused,and different pilot overheads including 1/7 and 2/7 areconsidered. The performance gain of the IMP scheme shownin Fig. 5 can also be observed in Fig. 6. Due to lower pilotcollision probability achieved by increasing pilot overhead, performance of the TSP scheme and the IMP scheme areboth improved obviously under 2/7 pilot overhead, and betterperformance can be provided by the IMP scheme at eitherBLER = 0.1 or BLER = 0.01.Fig. 7 shows the performance of the IMP scheme with 2pilots or 3 pilots under different pilot overhead. The IMPscheme with 3 pilots has similar performance to the casewith 2 pilots at BLER = 0.1, while at BLER = 0.01, the IMPscheme with 3 pilots can support more UEs at given SNR orrequire lower SNR at given number of UEs, especially forthe scenario with more access UEs. In other words, the IMPscheme configured with more pilots would have betterperformance at higher reliability region.
Fig. 6. Performance comparison between the TSP scheme and the IMPscheme with 2 pilots under different pilot overhead.Fig. 7. Performance of the IMP scheme with 2 pilots or 3 pilots underdifferent pilot overhead.Fig. 8. Comparison of average decoding attempts between the TSPscheme and the IMP scheme. ig. 8 shows the comparison of average decoding attemptsbetween the TSP scheme and the IMP scheme with 2 pilotsand 3 pilots under 1/7 pilot overhead at SNR=6dB. Thedecoding attempts of the IMP scheme with 2 pilots is about1.5 times that of the TSP scheme, and the decoding attemptsof the IMP scheme with 3 pilots would increase further. Thatis to say, the IMP scheme significantly improve the systemperformance at the cost of increasing the receiver’scomplexity. Because of lower accuracy of channel estimation,large residual error will be produced and transferred to nextround of blind MUD, not only affecting the detection of theremaining UEs, but also leading to some decoded UEs beingdetected again and the decoding attempts being increased.
Fig. 9. Performance comparison between the TSP scheme and the IMPscheme with 2 pilots under different pilot overhead.Fig. 10. Comparison of average decoding attempts between the TSPscheme and the IMP scheme.
B. Data-aided Channel Estimation Used in IC
In this subsection, reconstructed data aided channelestimation is used in IC to reduce the residual error. Theperformance comparison between the TSP scheme and theIMP scheme configured with 2 pilots under different pilotoverhead is shown in Fig. 9. It can be seen that, with data-aided channel estimation, the performances are significantlyimproved for both the TSP scheme and the IMP schemecompared to the results in Fig. 6. In addition, the IMPscheme with 2 pilots still has a performance gain of at least 2dB under different pilot overhead at BLER = 0.1 and highUE loading scenario. The performance gain is more obviousat BLER = 0.01. Due to the advantage of data-aided channelestimation, the performance improvement with 2/7 pilot overhead tends to be small, similar conclusion can also beobserved for the IMP scheme with 3 pilots, the results ofwhich are not shown in the figure.The comparison of average decoding attempts between theTSP scheme and the IMP scheme under 1/7 pilot overhead isshown in Fig. 10. With data-aided channel estimation, theaverage decoding attempts of the IMP scheme is reducedsignificantly and close to that of the TSP scheme.IV. C
ONCLUSION