Probe Selection for Over-the-Air Test in 5G Base Stations with Massive Multiple-Input Multiple-Output

Xiang Zhang,Shangbin Qiao,Mugen Peng,＊,Yong Li

1 China Academy of Information and Telecommunications Technology (CAICT),Beijing 100191,China

2 State Key Laboratory of Networking and Switching Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China

Abstract: In 5G systems,massive multiple-input multiple-output (MIMO) has been adopted in base stations (BSs) to improve spectral efficiency and coverage.The traditional conductive performance test techniques are challenging due to the unaffordable cost and high complexity when testing a large number of antennas.To solve this problem,the over-the-air (OTA) test has been presented,in which probe selection is the key to reduce the number of channel emulators and probes.In this paper,a novel artificial bee colony (ABC) algorithm is introduced to enhance the efficiency and accuracy of probe selection procedure.A sectoring-based multi-probe anechoic chamber (MPAC) is built to evaluate the throughput performance of massive MIMO equipped in 5G BS.In addition,link level simulation is carried out to evaluate the proposal’s performance gain under the commercial network assumptions,where the average throughput of three velocity is given with different SNR region.The results suggest that OTA chamber and multi-probe wall are available not only for 5G BSs,but also for user equipments (UEs) with end-to-end communication.

Keywords: anechoic chamber; massive MIMO base stations; over-the-air test; probe selection; performance evaluation

I.INTRODUCTION

To meet the increasing demand of high-speed and high-quality communication,many novel multi-antenna technologies have been utilized in mobile terminals and base stations (BSs).At the same time,the emergence of those new technologies will bring new requirements for performance test,which is an important part in equipment design and development stage.However,the traditional conductive test scheme has inherent disadvantages,e.g.,unrealistic performance evaluation caused by the omission of antenna pattern and high test complexity because of numerous cable connections.Moreover,multi-antenna devices are foreseen not to support radio frequency (RF) cable connections in the high frequency bands,which means that the conductive test is not available for mmWave 5G test.In view of this,a more accurate,more flexible and faster testing solution is needed to meet the new test requirements.Obviously,over-the-air (OTA) test,which is also called radiated test,is the promising solution to overcome these problems.

Because of the urgent need for radio performance test of multi-antenna devices,OTA test for multiple input multiple output (MIMO) has attracted more attentions from industry and academic community in recent years.Three key methods have been commercially adopted for the UE,which are the reverberation chamber (RC) method,the radiated two-stage (RTS) method and the multi-probe anechoic chamber based (MPAC) method,respectively.These three methods are only suitable for UE downlink throughput performance test in sub 6GHz.However,there are currently no widely accepted OTA performance solutions for massive MIMO base station (BS).According to [1],the MPAC method is seemed to be the most promising solution for massive MIMO performance test under clustered delay line (CDL) channel fading.

Extensive literatures have been published to discuss MPAC algorithms and results for long term evolution (LTE) UE.Power angle spread (PAS),which represents the channel spatial correlation,is the key parameter to optimize the weight of each probe in MPAC.Because the massive MIMO technique is applied in 5G BS,the PAS of BS is much smaller than that of UE,e.g.,5°.Therefore,the distance between two adjacent probes becomes smaller compared with UE MIMO OTA.In order to support the whole coverage of BS sector with limited probes,the location of multiple probes needs to be selected based on the cluster delay profile [1].In addition,the whole anechoic chamber setup is discussed and simulated in [2].One of the most important channel model in LTE stage,i.e.,spatial channel model extended (SCME),is reproduced in [3].Throughput is often adopted as the criterion for UE MIMO OTA test,but other metrics need be considered for massive MIMO BS,e.g.,beam peak distance,total variation distance of beam allocation distributions and fixed beam power loss [4].How to economically achieve the end-to-end throughput test has also been generally analyzed in [5].However,two essential problems needs to be carefully studied when we want to deploy the massive MIMO OTA chamber for BS throughput test.1) Because the length of 5G BS equipment is about 1m,the far field distance is larger than 30m,which is difficult to fulfill in the anechoic chamber.2) Plenty of channel emulators and probes are needed due to the multiple clusters and the expansion of test zone (TZ).Above two issues will dramatically increase the cost of whole massive MIMO OTA test system [6].Therefore,it is necessary to flexibly allocate the probe location according to the desired channel model.Several probe selection algorithms have been introduced in [7] for MIMO OTA test,which proves that the multishot algorithm is the most efficient one.However,whether this algorithm is suitable for massive MIMO techniques has not been verified.This paper focuses on how to generate the desired channel model in the large TZ with small number of probes for 5G massive MIMO BS.

The major contributions of this paper lie in the following aspects:A) The artificial bee colony (ABC) algorithm based on swarm intelligence is introduced to perform probe selection for massive MIMO OTA test.B) The classic multishot algorithm is compared with the ABC algorithm under the near field assumption.The simulation results show that the proposed ABC algorithm is accurate with low computational complexity.C) Link level simulation is carried out to evaluate the downlink throughput in two CDL channel models.The results show that the downlink throughput of 5G significantly outperforms LTE system.D) A large OTA chamber and multi-probe wall are built to evaluate the practical 5G end-to-end performance.Channel emulator instrument is also applied to generate doppler,spatial correlation of UE,power delay profile,and etc..

The paper is organized as follows.In Section II,the MPAC setup for massive MIMO and the system model of probe location optimization are given.Then,the novel ABC algorithm is proposed and compared with the multishot algorithm in Section III.In Section IV,the link level simulation is carried out to evaluate downlink throughput,and the practical OTA chamber architecture is introduced.Finally,we conclude the paper in Section V.

II.CHANNEL EMULATION IN ANECHOIC CHAMBER

2.1 MPAC setup for massive MIMO OTA test

In the MIMO OTA test of UE,eight cross polarized probes are located on a horizontal ring with equal degree and the device under test (DUT) is placed in the center.Because the antennas equipped in the UE is omnidirectional,downlink signal is radiated from all the probes with large angle spread.On the contrary,Due to the three dimensional (3D) beamforming and sectoring-based antennas are adopted in the 5G BS,a 3D sectoring MPAC[1],[4],[8] setup is utilized for the massive MIMO BS throughput performance test as illustrated in the figure 1.The blue rectangle in the center indicates the TZ and the BS panel should be located there in the test.Red dots distributed on the spherical surface,which is also called the probe wall in the rest of this paper,represent all potential locations of cross polarized probes.Several locations need be chosen to install probes according to the desired channel models.

Based on the coverage of 5G BS,the potential locations for probes are limited in the spherical surface with a horizontal range from -60°to 60°and vertical range from -21°to 21°.This area will also cover most of the angle of departure (AoD) in the CDL channel models [9].The potential locations of probes are evenly distributed in the sphere surface with fixed angle.The small angle of two adjacent probes will cause severe signal reflection and coupling,and enhance the computational complexity and the cost.If the two adjacent probes are far apart,small angle spread cannot be generated.Therefore,the intervals between two adjacent probes for the channel model in [9] need to be carefully optimized through extensive simulation.As illustrated in the figure 2,the interval between two adjacent rows is 3°,while the interval between two adjacent columns in the same row is 6°.Therefore,there are a total of 307 potential locations on the probe wall.

2.2 Test zone sampling method

The spatial correlation,which is a Fourier transform pair with the PAS,plays the key role in power allocation for the MIMO OTA probes.Because different PAS may result in the same spatial correlation,it is necessary to introduce some constraints,e.g.,angle spread (AS) and mean angle [10].However,some strict constrains may have no solutions of the objective function.Fortunately,there is no need to add the above restrictions when a good sampling method is utilized.The AS and mean angle are automatically matched with the target while the objective function still has a solution.

Fig.1.Anechoic chamber setup.

Fig.2.Potential locations of probes.

Some widely adopted sampling methods are studied in [11].Without losing generality,we choose deterministic grid algorithm in this paper as illustrated in the figure 3.The sampling points are located in the square grid of TZ,and each point is paired with the center one to calculate the target spatial correlation.

2.3 Problem formulation

In the MPAC MIMO OTA,there are two methods to realize the synthesis of propagation environment within the TZ,i.e.,prefaded signals synthesis (PFS) and plane wave synthesis (PWS) [12].The former one is more attractive in the practical system because it does not need phase calibration and requires less probes.In the PFS method,clusters are mapped to the probes based on the spatial correlation defined in the CDL channel models.

In order to get the spatial correlation distribution,the test zone is sampled into a plurality of discrete points.Ignoring the UE antenna pattern,the target spatial correlation between each sampling point and the center one can be expressed as

Fig.3.Sampling and pairing method.field measurement results in the outdoor NLoS scenario.

whereP(ϑ) andP(φ) are the vertical and horizontal power distribution,respectively.In this paper,the PAS is assumed to follow the truncated Laplacian distribution for each cluster in both direction.The truncated Laplacian distribution for the angle∈is written as [13]

where:

·∈denotes either vertical angleϑor horizontal angleφ.

·QLis scaling constant ensuring ∮Pd(Ω Ω=) 1.

·σis the standard deviation.

·∈0represents the mean angle of∈.The value range ofφandϑare different,whereφc onfines within [-π+φ0,π+φ0] andϑis within

Combining (1) and (2),a complete 3D spatial correlation formula between two points can be obtained as

The above equation represents the spatial correlation between two points of the single cluster.However,it is not possible to select probes based on the individual cluster optimization because of limitation of probe number.Therefore,the probe selection algorithm needs consider the total spatial correlation of multiple clusters.Multiple clusters emulations are discussed in [14] and the spatial correlation of multiple clusters can be summarized as

wherep(n)is the normalized power of thenthcluster andρnis the spatial correlation of the clusternwhich is calculated by (4).

In order to match the target spatial correlation in (5),different weights are assigned to theKprobes.The emulated spatial correlation ofKprobes between two points can expressed as

Assuming that there areMsampling points in the TZ,and the error between the emulated and target spatial correlation can be minimized as

The probe positions and weights need to be simultaneously optimized in (7),which is a complicated NP-hard problem.In fact,when the probe positions are fixed and only weights need to be optimized,it will degenerate into convex optimization problem that can be effectively solved [10] [13].There areCSKcombinations in total ifKprobes are selected fromSpotential positions.Therefore,this is a combinatorial optimization problem,in which the merits of each combination can be obtained by the convex optimization process.

III.ALGORITHM FOR FLEXIBLE PROBE SELECTION

As described in the section II,the probe selection can be viewed as a combinatorial optimization problem.The global optimal solution can be obtained by traversing all combinations.However,due to the large number of potential positions in the massive MIMO probe wall,a large amount of combinations makes this method impractical.

In order to efficiently select the probes,a novel algorithm called multishot is proposed and analyzed to solve combinatorial optimization problems [7].Some hybrid approaches may be introduced to further reduce the computation complexity and enhance the accuracy,e.g.,artificial neural networks algorithm,genetic algorithm (GA),particle swarm optimization (PSO) algorithm,and ABC algorithm.Without losing generality,the ABC algorithm is applied in this paper for the probe selection because of its fast convergence and strong global search ability.

3.1 Multishot algorithm

The basic idea of multishot algorithm relies on it that probes with smaller weights contribute less when synthesizing the target spatial correlation.Therefore,probes with small weights can be removed.The calculation procedure of the multishot algorithm is illustrated in the figure 4.

Multishot is a greedy algorithm,which follows the problem solving heuristic of making the locally optimal choice at each stage with the intent of finding a global optimum.However,a local optimal solution may be obtained if the greedy strategy is not appropriate.One probe may have different weights in different combinations,which means that the probe may not be important in some combinations but contribute a lot in other combinations.Therefore,it is entirely possible to remove some probes in the global optimal set during executing the multishot algorithm.

3.2 Artificial bee colony algorithm

Fig.4.Flowchat of multishot algorithm.

Fig.5.Flowchat of ABC algorithm.

The ABC is a novel optimization algorithm inspired by the honey collecting behavior of bee colony.Bees perform different activities according to their respective responsibilities and modify their behaviors through communication and information sharing.As a result,the best honey source could be found in the end.Some important concepts in the ABC algorithm are listed as follows.

3.2.1 honey source

Each honey source represents a feasible solution,which is a combination ofKprobes for the probe selection problem.When the positions of probes are fixed,(7) is rewritten as

Each honey source can obtain its optimal value according to (8) and the optimal value reflects the quality of honey source.

3.2.2 leading bees

There is a one-to-one correspondence between leading bees and honey sources.The leading bees search the adjacent source in each iteration and replace the previous source if founding a better one.

3.2.3 following bees

The number of following bees and leading bees is the same.Before starting a new round search,each following bee will select one leading bee based on the quality of the best honey source and search along with the selected leading bee in the next round.

3.2.4 scouting bees

Leading bees and following bees are constantly looking for the better honey source near to the previous source.If the previous honey source is still not updated after a certain number of iterations,then it should be removed and randomly replaced by a new honey source.These new honey sources correspond to the scouting bees.

In general,the leading bees are used to maintain a good solution,the following bees are used to accelerate the convergence and the scouting bees are used to jump out of the local optimal solutions.The procedure of ABC algorithm,which is applied in probe selection procedure,is illustrated in the figure 5.

3.3 Simulation comparison

The parameter assumptions in the simulation are shown in the table 1.

In order to improve the experimental efficiency,the sampling intervals of TZ are set 0.02m and 0.01m when the number of available probes are 8 and 4,respectively.In the CDL-C channel model [9],there are 24 clusters but it is impossible to reproduce all the clusters because of the limited number of probes.To avoid the unnecessary computation complexity,we only consider the clusters whose powers are higher than 25dB below the maximum cluster power,and the clusters,whose AoDs are out of the probe wall,are removed.

For the multishot algorithm,only one probe is removed from each iteration,while for the ABC algorithm,the maximum number of allowed non-update iterations for global best honey source is 20 generations.Maximum non-updated iteration number for each local honey source is set to 10.

Equation (7) represents the length of error vector between emulated and target spatial correlation,which reflects the overall error of all sampling points in the TZ.The normalized power of all probes after performing the first optimization is illustrated in the figure 6.It shows that it is superfluous to implement the channel model with 307 probes.In fact,20～30 probes are enough to generate the CDL-C channel model if the potential positions of probes is defined in the figure 2.This view is further demonstrated by the simulation shown in the figure 7.In the early stage of multishot algorithm,the correlation error does not increase with the number of probes decreasing,but the error will grow when the number of probes is less than 18.

The simulation results show that the multishot algorithm can almost get the global optimal solution when there are enough probes.However,the performance of this algorithm needs to be improved when the number of probes are limited.In addition,much meaningless selection procedure is operated in theearly stage,which could also be optimized.

Table I.Parameter settings.

Fig.6.Probe weight after the first optimization.

Fig.7.Error vs.the number of probes.

In order to obtain a good quality of first generation honey source and reduce the search area,a convex optimization process involving all probes needs to be performed before the ABC algorithm executing.After removing the probes with small power weights,the searching area will be dramatically reduced and the efficiency of the ABC algorithm can be improved.

The error between emulated and target spatial correlation of sampling points in the TZ is shown in the figure 8 and figure 9.The optimization results of these two algorithms are almost the same when the number of available probes is 8.The total error calculated by (7) is 7.52 for multishot algorithm and 7.42 for ABC algorithm.However,the ABC algorithm is obviously outperforms the multishot algorithm when the available probes are 4 as illustrated in the figure 9.

Fig.8.Spatial correlation error with 8 probes.

Fig.9.Spatial correlation error with 4 probes.

IV.LINK LEVEL SIMULATION AND PRACTICAL SYSTEM

Link level simulation is carried out to evaluate 5G end-to-end downlink throughput performance under the CDL channel fading.Based on the China mobile’s requirement,there are 64 transmission and reception RF ports deployed in the BS,and each RF port is connected to three antenna elements with the same polarization angle,so the total number of antenna elements is 192.The spacing between two adjacent horizontal and vertical elements is 0.5λand 0.7λ,respectively.There are four cross-polarized antennas in the 5G prototype UE considering the size and cost.The spacing of two adjacent antennas with same polarization is 0.5λ,and the UE antennas are orthogonally polarized at the angle of 0°and 90°from the vertical.Therefore,only four data streams are supported simultaneously.The antenna gain of gNB and UE is define in [9],and the UE is moving to the BS along the boresight direction.Both low and high velocity is considered.According to the commercial deployment requirement,the central frequency of the simulation is 3.5GHz and the bandwidth is 100MHz.Normal CP and frame structure option 2 are adopted.The downlink data channel cannot occupy the resource element of control channels.The precoding matrix index,channel quality indicator and rank indicator are feedback to the BS,and the BS can choose whether adopt these information for beamforming.Adaptive modulation and hybrid automatic repeat request schemes are utilized to decrease the impact of fast fading.High order modulation,i.e.,256QAM,is applied when the signal to interference plus noise ratio (SINR) is high enough.The low density parity check code (LDPC) replaces Turbo code in 5G for its high coding efficiency when size of data block is large.Channel state information-reference signal (CSI-RS) and demodulation reference signal (DMRS) are introduced for channel quality measurement and detection.Without loss of generality,CDL-C and CDL-D channel models are applied to evaluate the downlinkperformance under none line of sight (NLOS) and line of sight (LOS) scenarios respectively.The detailed parameters are listed in the table 2.

Table II.Parameter settings.

Fig.10.The downlink throughput for CDL-C model.

Fig.11.The downlink throughput for CDL-D model.

Fig.12.The BS beam pattern for different distance.

Fig.13.The picture of probe wall.

As depicted in the figure 10,the average downlink throughput of CDL-C channel model is given for different SNR region and UE velocity.When the UE speed is 3km/h and the SNR is 32dB,4 data streams can be supported and the peak rate of 1.6Gbps can be almost achieved.However,the data throughput of 120km/h is nearly the half of 3km/h in high SNR region,because the multiplexing transmission is more sensitive to the high doppler.At most two data streams can be scheduled in the 120km/h even there is no interference and noise.The throughput is reduced as the SNR deceases,but the throughput is always higher than 180Mbps,which significantly outperforms LTE commercial networks.The LOS performance of CDL-D channel model is given by the figure 11.Only 3 data streams can be supported because the spatial correlation of CDL-D is much higher than that of CDL-C.When the SNR is 24dB with 3km/h,the rank is reduced to 2 which is achieved by the diversity of cross polarization.The performance gap between CDL-D and CDL-C becomes smaller in low SNR region.Due to the high doppler in 120km/h,the peak rate is less than 500Mbps in CDL-D model.More effective detection algorithms need to be considered in the future.

The length of gNB for 5G sub 6GHz is larger than 1m due to multiple antennas are deployed.The far field distance equation is given bywhereDrepresents the length of BS.The far field distance is larger than 30m,which is not practical in the anechoic chamber.As illustrated in the figure 12,we simulate the beam pattern for different distance between the probe wall and BS based on the antenna element architectures of 5G BS in 3.5GHz.The 3dB beam pattern is almost the same when the distance is larger than 3m.Because the multi-user performance is not considered,the impact of null point can be ignored.Therefore,3.5m is chosen for the radius of probe wall to keep the balance between the cost and performance,and the size of the anechoic chamber is set to 10m×10m×10m.The picture of the probe wall is given by the figure 13.

As analyzed in the section III,ABC algorithm is chosen for the selection of probe location.There are total 18 cross polarized probes applied to generate the AoD,ZoD,ASD,and ZSD of CDL-C and CDL-D channel models.

The positions of clusters defined by the AoD and ZoD are given by the figure 14.It is noted that the clusters,which are out of the beam coverage or have small power,have been canceled.The optimized probe locations are around the clusters to generate the desired PAS.Each probe may contribute to 3 or 4 clusters with multiple weights.The BS downlink beam signal is received by multiple probes,and then transmitted to the channel emulators by the cables.Doppler,power delay profile in each cluster and the channel spatial correlation in the receiver are added by the channel emulators.Because all the transmission links are bidirectional,the proposed end-to-end performance test system can support both downlink and uplink.Thanks to the characteristic of channel reciprocity in TDD system,uplink channel estimation is adopted for downlink beamforming.Therefore,the capability of this massive MIMO OTA system is much stronger than that of current commercial UE MIMO OTA system.

V.CONCLUSIONS

In order to meet the test requirement of 5G massive MIMO BS,a sectoring-based MPAC setup is proposed to evaluate the throughput performance.Considering the cost and efficiency,a novel ABC algorithm is utilized to select the probe location in the massive MIMO OTA chamber.The simulation results show that ABC algorithm outperforms the classic multishot algorithm significantly when the number of probes is limited.Accurate link level simulation is carried out to evaluate the performance under the commercial network assumptions.The downlink throughput of 5G is much higher than that of LTE in all SNR region.Moreover,the probe wall and anechoic chamber are built not only for 5G BSs,but also for UEs with end-to-end communication.Moreover,previous link level simulation results can be adopted as the criterion to check transmission and detection algorithms.In the future work,different types of 5G BS and UE will be checked and analyzed in our massive MIMO OTA chamber.

Fig.14.The location of probes and clusters.

ACKNOWLEDGEMENTS

This work was supported by the State Major Science and Technology Special Projects under Grant No.2018ZX03001028-003.