ieice communications express, vol.8, no.10, 404 comparison ... · is a typical usage scenario in...

Comparison of angle-of-arrival characteristics at2.4GHz and 60GHz bands

Takuto Kurose, Satoru Kishimoto, and Minseok Kima)

Graduate School of Science and Technology, Niigata University,

8050 Ikarashi 2-no-cho, Nishi-ku, Niigata 950–2181, Japan

a) [email protected]

Abstract: For seamless communication in millimeter-wave (mm-wave)

transmission systems, the robustness against link blockage and user mobility

should be guaranteed. Cooperative joint network design over conventional

microwave bands and mm-wave bands is essential in future mm-wave

WLANs (e.g., IEEE 802.11ay) and 5G cellular networks, and hence under-

standing the discrepancy between the propagation properties at those fre-

quency bands is crucial. In this letter, the angle-of-arrival characteristics at

mm-wave band (60GHz) and microwave band (2.4GHz) in indoor environ-

ments are presented. From the measurement results, it was seen that the line-

of-sight and first-order reflected paths agree well each other, but diffraction

and scattering are observed only at microwave band. It was also shown that

the angular spreads at mm-wave band was about 25 degrees smaller than

those at microwave band.

Keywords: fast session transfer, millimeter wave, microwave, channel

sounding, angle-of-arrival, angle spread, antenna array

Classification: Antennas and Propagation

References

[1] “Channel models for 60GHz WLAN systems,” IEEE Document 802.11-09/0334r8, May 2010.

[2] “Channel models for IEEE 802.11ay,” IEEE Document 802.11-15/1150r2,Sept. 2015.

[3] A. Patra, L. Simić, and M. Petrova, “Design and experimental evaluation of a2.4GHz-AoA-enhanced beamsteering algorithm for IEEE 802.11ad mm-waveWLANs,” IEEE 18th International Symposium on A World of Wireless,Mobile and Multimedia Networks (WoWMoM), July 2017. DOI:10.1109/WoWMoM.2017.7974290

[4] S. Sur, I. Pefkianakis, X. Zhang, and K. Kim, “WiFi-assisted 60GHz wirelessnetworks,” The 23rd Annual International Conference on Mobile Computingand Networking, Snowbird, Utah, United States, Oct. 2017. DOI:10.1145/3117811.3117817

[5] C. Gustafson, F. Tufvesson, S. Wyne, K. Haneda, and A. F. Molisch,“Directional analysis of measured 60GHz indoor radio channels using SAGE,”IEEE 73rd Vehicular Technology Conference (VTC Spring), Yokohama, Japan,May 2011. DOI:10.1109/VETECS.2011.5956639

[6] J. Medbo, N. Seifi, and H. Asplund, “Frequency dependency of measured

© IEICE 2019DOI: 10.1587/comex.2019XBL0093Received June 10, 2019Accepted June 28, 2019Publicized July 16, 2019Copyedited October 1, 2019

404

IEICE Communications Express, Vol.8, No.10, 404–409

http://dx.doi.org/10.1109/WoWMoM.2017.7974290





http://dx.doi.org/10.1145/3117811.3117817

http://dx.doi.org/10.1145/3117811.3117817

http://dx.doi.org/10.1145/3117811.3117817

http://dx.doi.org/10.1145/3117811.3117817

http://dx.doi.org/10.1109/VETECS.2011.5956639




highly resolved directional propagation channel characteristics,” The 10thEuropean Conference on Antennas and Propagation (EuCAP), Oulu, Finland,May 2016.

[7] J. Medbo, H. Asplund, and J.-E. Berg, “60GHz channel directional character-ization using extreme size virtual antenna array,” IEEE 26th AnnualInternational Symposium on Personal, Indoor, and Mobile Radio Communi-cations (PIMRC), Hong Kong, China, Sept. 2015. DOI:10.1109/PIMRC.2015.7343290

[8] S. Kishimoto, M. Kim, D. He, and K. Guan, “Scattering process identificationand cluster analysis for millimeter-wave indoor channel model,” 2018International Symposium on Antennas and Propagation (ISAP 2018), Oct.2018.

[9] T. Min, K. Saito, and J. Takada, “Development of directional channel sounderusing USRP and GNU radio,” ASEAN Eng. J., vol. 7, no. 1, 2017.

[10] C. L. Dolph, “A current distribution for broadside arrays which optimizes therelationship between beam width and side-lobe level,” Proc. IRE, vol. 34,no. 6, pp. 335–348, June 1946. DOI:10.1109/JRPROC.1946.225956

1 Introduction

Recently, the demand for ultra high-speed wireless data transfer for various new

applications such as ultra-high definition (4K/8K) and virtual/augmented reality

(VR/AR) technologies is increasing. The technical standard for multi-gigabit

WLANs at 60GHz millimeter-wave (mm-wave) band has developed in IEEE

802.11ad [1] and the advanced version is currently being delopved in IEEE

802.11ay [2] which can support up to 30Gbps throughput. However, the prop-

agation loss at mm-wave bands is significantly large and the attenuation by

diffraction and penetration is also very large. In this regard, the functionality of

fast session transfer (FST) is seriously considered in IEEE 802.11ad [1] to

seamlessly switch to Wi-Fi (2.4/5GHz) when the mm-wave link becomes un-

available due to blockage or beam misalignment. However, existing FST tech-

niques need time-consuming sector sweep and power-consuming mm-wave chan-

nel monitoring, thus various multiband techniques have been studied for more

efficient FST [3, 4]. In order to design multiband WLANs the discrepancy between

the propagation characteristics of different frequency bands should be investigated.

Especially, the angular properties such as power spectra of angle-of-arrival (AoA)

and angle-of-departure (AoD), are important to apply spatial transmission tech-

niques such as beamforming and MIMO (multiple-input-multiple-output).

In order to compare the angular channel characteristics at different frequency

bands, they should be measured by using the identical measurement conditions.

Several super-resolution parameter estimation methods such as SAGE and RIMAX

[5] which can extract the multi-path components (MPCs) where the entire response

of the measurement system including the antennas can be removed. These para-

metric methods are based on the assumption that the radio channels can be

decomposed by a set of discrete plane waves and diffuse scattering. However,

since we may have ambiguous decomposition of the channel components due to the

signal processing limitation in treating the diffuse scattering, the angular power


405


http://dx.doi.org/10.1109/PIMRC.2015.7343290




http://dx.doi.org/10.1109/JRPROC.1946.225956




spectrum (APS) is more appropriate way for the comparison purpose [6, 7]. In this

letter, we compared the AoA properties at two different frequency bands from the

APS where angular scanning of high-gain horn antenna and virtual cylindrical array

were employed for mm-wave band and microwave band, respectively. It should be

noted that the half-power beamwidth (HPBW) of the virtual cylindrical array was

designed equal to that of the horn antennas as much as possible to achieve identical

angular resolution. The observation from the measurement results and discussion

on the discrepancy of the propagation mechanisms and angular characteristics are

presented.

2 APS measurement methods

2.1 Millimeter wave band

For mm-wave band, a full polarimetric 2 � 2 MIMO channel sounder at the center

frequency exactly of 58.5GHz was used [8]. Rotating highly directive horn

antennas with a gain of 24 dBi (HPBW of 12 degrees) at both transmitter (Tx)

and receiver (Rx), the 5-dimensional channel transfer functions (CTFs) of 256 sub-

carriers over 400MHz bandwidth as

Hqpðfk; �0i0 ; �

0j0 ; �i; �jÞ ð1Þ

were measured by transmitting an unmodulated multitone signal where the sub-

script p; q 2 f�; �g denote polarization for Tx and Rx antennas, respectively, and � 0i0and �0

j0 , and �i and �j indicate the i0-th pointing co-elevation (zenith) and the j0-thpointing azimuth angle at Tx, and the i-th pointing co-elevation and the j-th

pointing azimuth angle at Rx, respectively. The double-directional angle delay

power spectrum (DDADPS) is given by

Pqpðl; i0; j0; i; jÞ ¼ jhqpð�l; � 0i0 ; �0j0 ; �i; �jÞj2 ð2Þ

where the channel impulse response hqp is obtained by inverse Fourier transform of

(1). From DDADPS, the polarization combined APS is synthesized by

APSð�i; �jÞ ¼ 1

2

Xp,q2f�;�g

Xl;i0;j0

Pqpðl; i0; j0; i; jÞ ð3Þ

where the angle sampling interval is typically 12 degrees.

2.2 Microwave band

For microwave band, a software radio based narrowband channel sounder [9]

having 400 kHz bandwidths at center frequency exactly of 2.425GHz was used. In

Fig. 1(a), the cylindrical antenna array structure having half-wavelength element

spacing is shown where the APS was synthesized by beamforming with the patch

antenna element of 7.3 dBi gain. The response function at the position ðm; nÞ form ¼ 1; � � � ; M, and n ¼ �N; � � � ; N can be expressed as

An;mð�; �Þ ¼ Aucan;mð�; �ÞAula

n ð�ÞUmð�; �Þ ð4Þwhere Auca

m and Aulam denote the circular and linear array responses, respectively,

which are expressed as© IEICE 2019DOI: 10.1587/comex.2019XBL0093Received June 10, 2019Accepted June 28, 2019Publicized July 16, 2019Copyedited October 1, 2019

406


Aucam ð�; �Þ ¼ exp j

2�

�r sin � cos � � 2�

Mðm � 1Þ

� �� ; ð5Þ

Aulan ð�Þ ¼ Gn expð j2�dn cos �Þ ð6Þ

where r, λ, M and N denote the radius, wavelength, and the numbers of antennas in

the circular and linear arrays, respectively. Umð�; �Þ and Gn denote the patch

element pattern and the Dolph-Chebyshev coefficients [10], which result in side-

lobe reduction in horizontal and vertical planes, respectively. As shown in

Fig. 1(b), it is seen that the cylindrical array beampatterns synthesized by the patch

elements (M ¼ 25 and N ¼ 7) are well matched to the horn antenna patterns. The

APSð�; �Þ for i-th pointing co-elevation and the j-th pointing azimuth angle at Rx

is obtained by the beamforming as

APSð�i; �jÞ ¼ wHð�i; �jÞRxxwð�i; �jÞ ð7Þwhere the correlation matrix Rxx ¼ E½xðtÞxHðtÞ� 2 C

Mð2Nþ1Þ�Mð2Nþ1Þ defining

xðtÞ 2 CMð2Nþ1Þ by the input signal vector, and the element of the weight vector

½wð�i; �jÞ�MðNþnÞþm ¼ An;mð�i; �jÞ. Here, (7) is calculated every 12 degrees as in

(3).

2.3 Angle spread

The angular channel characteristics at different frequency bands are compared using

the azimuth angle spread calculated with the azimuth power spectrum as

AzPSð�jÞ ¼P

iAPSð�i; �jÞ. The azimuth angle spread distribution is usually

modeled as Wrapped-Gaussian distribution. The angle spread is calculated by

� ¼ min�

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXjð�jð�ÞÞ2 � AzPSð�jÞX

jAzPSð�jÞ

�X

j�jð�Þ � AzPSð�jÞX

jAzPSð�jÞ

0@

1A

2vuuut ; ð8Þ

�jð�Þ ¼ modð�j þ �; 360�Þ ð9Þwhere modðÞ denotes the modular operator, and Δ indicates an offset angle.

(a) Virtual cylindrical array. (b) Comparison between cylindricalarray synthetic beampatterns andhorn antenna radiation patterns

Fig. 1. Scattering process identification results.


407


3 Measurement results

3.1 Measurement scenarios

We conducted the measurement campaign at a conference room environment which

is a typical usage scenario in IEEE 802.11ad and 802.11ay. The Tx as an access

point (AP) was set on the television at a height of approximately 2.1m close to the

wall where the antenna radiation pattern covers the front side of the wall. The

channel responses were measured at the five Rx positions (denoted by Rx1–Rx5)

where the Rx was assumed to be a station (STA), e.g., a laptop PC. The STA

antenna was located from 3.12 to 5.65m away from the AP at the height of 0.9m

from the floor (15 cm from the table). As can be seen, the LoS between Tx and Rx

was available in all Rx positions. This refers to the setup of the STA-AP conference

room sub-scenario in [1].

3.2 Results and observation

Fig. 2 shows the measured APS at Rx5 as an example where the ray tracing

simulation result is also presented with circular markers in Fig. 2(a) as a reference.

In addition, Fig. 3 illustrates the identified propagation mechanisms which were

confirmed by the ray tracing results. It is noted that the APS is normalized by the

maximum power and the values larger than −25 dB are only considered. It is seen

that the LoS (#1) and first-order reflected paths (#2, #3 and #4) are well observed at

both frequency bands. However, the higher-order reflected paths (with diffraction)

and diffuse scattering around window frames and metallic chairs (e.g., #5) are

observed only at microwave band. The interpretation is as following. At microwave

band, it is well known that the scattering objects such as furniture mainly acts as

scatterers rather than reflectors due to their size comparable to the wavelength, but

the specular reflection is mainly generated by the large object such as walls, ceiling

and floors. However, at mm-wave band, it can be expected that even small objects

may behave as specular reflector due to the very short wavelength, and hence

diffuse scattering was reduced.

In order to quantify the above mentioned discrepancy, angle spread which is

one of widely accepted parameters is employed. Fig. 4 shows the angle spread at all

Rx positions which are also indicated in the same figure. As shown in this figure,

the correlation of angle spread between the two frequencies are not clearly shown

(a) Microwave

#1 #2#3 #4 #5

(b) Mm-wave

#1 #2#3 #4

Fig. 2. Measured APS at Rx5 ((a) microwave and (b) mm-wave)


408


because the number of measurements is limited and hence the environment-de-

pendency could not be perfectly eliminated. However, it can be seen that the

average angle spread �mmwave at mm-wave band is significantly smaller (about

25 degrees) than ��wave at microwave band. That can be explained as the most

propagation mechanisms are specular reflection and there is little contribution of

the higher-order reflection (with diffraction) and diffuse scattering at mm-wave

band.

4 Conclusion

In this letter, we conducted dual-frequency channel measurements at mm-wave

band (60GHz) and microwave band (2.4GHz) in an indoor environment to

evaluate the AoA characteristics. From the measurement results, it is seen that

the LoS and first-order reflected paths agree well each other, but diffraction and

diffuse scattering at mm-wave band were not so significant. The angular spread

which is calculated by assuming Wrapped-Gaussian distribution was highly envi-

ronment-dependent, and the correlation between the two frequencies was not

clearly shown. However, it was observed that the average angle spread at mm-

wave band is significantly smaller (about 25 degrees) than that at microwave band.

These results should be utilized for cooperative joint network design using micro-

wave and mm-wave bands for future wireless systems.

Acknowledgments

This research and development work was supported by the MIC/SCOPE

#195004002.

Fig. 4. Angle spreadsFig. 3. Identified mechnisms at Rx5


409


Performance of synchronousCDMA for the PLC-basedremote multi-machine control

Mitsuru Hasegawa1a), Kentaro Kobayashi2, Hiraku Okada2,and Masaaki Katayama21 Dept. of Information and Communication Engineering, Nagoya University,

Furo-cho, Chikusa-ku, Nagoya-shi, Aichi 466–8603, Japan2 Institute of Materials and Systems for Sustainability, Nagoya University,

Furo-cho, Chikusa-ku, Nagoya-shi, Aichi 466–8603, Japan


Abstract: This paper considers a multi-machine control system using

power line communication (PLC). The signal-to-noise ratio (SNR) of PLC

channels has cyclostationary features synchronous to the mains voltage. As

a promising candidate of the multiple access scheme for the system, this

paper proposes a synchronous code division multiple access (SCDMA)

scheme that uses mains voltage as its system clock. By using orthogonal

codes, the communication performance of each code-channel is equalized,

and the worst-case performance is improved.

Keywords: PLC, synchronous code division multiple access (SCDMA),

multiple machine control, cyclostationary channel

Classification: Wireless Communication Technologies

References

[1] X. Zhang, Q. Han, and X. Yu, “Survey on recent advances in networked controlsystems,” IEEE Trans. Ind. Informat., vol. 12, no. 5, pp. 1740–1752, Oct. 2016.DOI:10.1109/TII.2015.2506545

[2] F. Passerini and A. M. Tonello, “Smart grid monitoring using power linemodems: Effect of anomalies on signal propagation,” IEEE Access, vol. 7,pp. 27302–27312, 2019. DOI:10.1109/ACCESS.2019.2901861

[3] Y. Ishigaki, Y. Kimura, I. Matsusue, H. Miyoshi, and K. Yamagishi, “Optimalenergy management system for isolated micro grids,” SEI Technical Review,pp. 73–78, 04 2014.

[4] S. Sawada, K. Kobayashi, H. Okada, and M. Katayama, “Selective transmissionof control information based on channel periodicity in plc-based multiple-machine control,” 2017 IEEE International Symposium on Power LineCommunications and Its Applications (ISPLC), pp. 1–6, Apr. 2017. DOI:10.1109/ISPLC.2017.7897109

[5] M. Katayama, T. Yamazato, and H. Okada, “A mathematical model of noise innarrowband power line communication systems,” IEEE J. Sel. Areas Commun.,vol. 24, no. 7, pp. 1267–1276, July 2006. DOI:10.1109/JSAC.2006.874408

[6] A. Kawaguchi, H. Okada, T. Yamazato, and M. Katayama, “Correlations ofnoise waveforms at different outlets of power-line network,” IEICE Trans.Fundamentals, vol. 90-A, no. 11, pp. 851–860, Nov. 2007.


410


http://dx.doi.org/10.1109/TII.2015.2506545




http://dx.doi.org/10.1109/ACCESS.2019.2901861




http://dx.doi.org/10.1109/ISPLC.2017.7897109





http://dx.doi.org/10.1109/JSAC.2006.874408




1 Introduction

The demand for the remote control for factory automation and smart grids is on the

increase [1]. As means of communication for indoor remote control, power line

communication (PLC), using the existing power lines as communication media,

has attracted considerable attention [2]. This paper considers narrow band-PLC

(NB-PLC) that is suitable for long-distance communication between the controller

and multiple machines in large factories and buildings.

In feedback control, communication is repeated periodically. The period of

communication is often about 10 [ms] in smart grid equipment [3]. This cycle is

close to that of the absolute value of the mains voltage flowing in the power lines.

Therefore, the authors considered the application of time division multiple access

(TDMA) which uses the strong mains voltage as a clock for system synchronization

[4].

Since the impedance and the noise of the electrical equipment connected to the

lines vary synchronously with (the absolute value of ) the mains voltage, the signal-

to-noise power ratio (SNR) also fluctuates according to the cyclic mains voltage

[5]. As a result, the communication quality of TDMA time slot assigned to each

machine varies, and when multiple machines work together such as in factories, the

poor quality machines dominate the total quality of the entire system. If we use

frequency division multiple access (FDMA) instead of TDMA, this variation of

SNR per channel still may not be mitigated as PLC noise is non-white [5].

This study thus proposes synchronous code division multiple access (SCDMA)

with orthogonal codes using the stable mains voltage with high SNR as a clock.

By performing communication of each machine at the same time and frequency

using code channel, the communication quality of all channels is made to be

uniform.

2 System overview

2.1 1:M feedback control system using NB-PLC

Fig. 1(a) shows the NB-PLC based 1:M multiple machines control system dis-

cussed in this paper. The feedback controller controls M machines via power lines

at every control cycle of TC seconds. In Fig. 1(a), xm½i� is the state information of

the m-th (m ¼ 0; 1; 2; � � � ; M � 1) machine in the i-th (i ¼ 0; 1; 2; � � �) control periodtransmitted in an NP-bits packet. The receiver output corresponding to xm½i� is

xm½i�. Based on xm½i�, the controller calculates the control command um½i�, andsend it back to the machine by an NP-bits packet. The receiver output at the m-th

machine corresponding to um½i� is um½i�.

2.2 NB-PLC channel model

In PLC systems, the receivers connected to the same power lines often observe the

same time variation of SNR [6]. Therefore, in this paper, we assume that the noise

voltage waveform nðtÞ in the signal band is common to all receivers1. Furthermore,

it is assumed that nðtÞ is cyclostationary colored Gaussian noise with a cycle

duration TN ¼ TAC=2 (TAC is the main voltage cycle duration). The power spectral

1Precisely, nðtÞ is the noise voltage normalized by the received signal voltage


411


density, mean, and variance of nðtÞ are respectively given by the following

equations [5]

SnðfÞ ¼ a

2expð�ajfjÞ; ð1Þ

E½nðtÞ� ¼ 0; ð2Þ

E½n2ðtÞ� ¼ �2ðtÞ ¼XL�1‘¼0

A‘jsinð2�t=TAC þ �‘Þjn‘ : ð3Þ

2.3 TDMA for NB-PLC

For comparison, we first describe the system using TDMA [4]. As shown in

Fig. 1(b), the control period is divided into 2M slots.

The binary phase shift keying (BPSK) modulated signal in the μ-th slot

(0 � � < 2M ) of the i-th control cycle is expressed as

s½T�i;� ðtÞ ¼ffiffiffiffiffiffi2P

p XNP�1

p¼0Re½bi;�½p�hcðt � pTS � T ½T �

i;� Þ expðj!ctÞ�; ð4Þ

T ½T �i;� ¼ �T ½T�

P � iTC; ð5Þwhere P is the received carrier power, hcðtÞ is a pulse of the symbol duration TS, and

bi;�½p� 2 f�1g is the p-th bit of the packet transmitting the status information

x�=2½i� for an even μ or the control command uð�þ1Þ=2½i� if μ is odd.

The receiver demodulates the signal arrived with noise nðtÞ by integrating for

the bit duration Tb to obtain the sample corresponding bi;�½p� to each bit expressed

as

r½T�i;� ½p� ¼1

Tb

Z ðpþ1ÞTbþT ½T �i;�

pTbþT ½T �i;�

ðs½T�i;� ðtÞ þ nðtÞÞdt ¼ffiffiffiffiffiffi2P

pbi;�½p� þ n½T�i;� ½p�: ð6Þ

(a) The 1:M multi-machines control system

(b) Slot assignment (TDMA) (c) Slot assignment (SCDMA)

Fig. 1. System model.


412


The mean and variance of this Gaussian random sample are given by

E½r½T�i;� ½p�� ¼ffiffiffiffiffiffi2P

pbi;�½p�; ð7Þ

E½ðr½T�i;� ½p�Þ2� ¼1

Tb

Z ðpþ1ÞTbþT ½T �i;�

pTbþT ½T �i;�

�2ðtÞdt ¼ �2½T�i;� ½p�: ð8Þ

Based on this sample r½T�i;� ½p�, the receiver makes the decision.

2.4 Synchronous CDMA for NB-PLC

SCDMA multiple signals are transmitted simultaneously. Therefore, the control

period is divided into two equal length slots, as shown in Fig. 1(c).

In the SCDMA system, the data bit stream is first spread by an orthogonal

spreading code

wm ¼ ðwm½0�; wm½1�; � � � ; wm½M � 1�Þ: ð9ÞThen, each chip is interleaved to mitigate the influence of bursty noise. The signals

from and to the m-th machine in the i-th control cycle are thus expressed as s½C�i;2mðtÞand s½C�i;2mþ1ðtÞ, respectively, with s½C�i;� ðtÞ is defined as

s½C�i;� ðtÞ ¼ffiffiffiffiffiffi2P

p XMNP�1

q¼0Re½ci;�½q�hcðt � qTS � T ½C�

i;� Þ expðj!ctÞ�; ð10Þ

T ½C�i;� ¼ ð� � 2b�=2cÞT ½C�

P þ iTC; ð11Þwhere ci;�½q� 2 f�1g is an interleaved chip data given by

ci;�½q� ¼ bi;�½q � NPbq=NPc�wb�=2c½bq0=NPc�; ð12Þwhere b�c is the floor function.

The receiver demodulates the signals s½C�i;� ðtÞ arrived with noise nðtÞ and the

integration with the chip length TS, and obtains the received samples corresponding

to each chip expressed as

r½C�i;� ½q� ¼1

TS

Z ðqþ1ÞTSþT ½C�i;u

qTSþT ½C�i;u

XM�1

m¼0s½C�i;2mþð��2b�=2cÞðtÞ þ nðtÞ

!dt

¼ffiffiffiffiffiffi2P

p XM�1

m¼0ci;2mþð��2b�=2cÞ½q� þ n½C�i;� ½q�: ð13Þ

Thus, its mean and variance are respectively given by

E½r½C�i;� ½q�� ¼ffiffiffiffiffiffi2P

p XM�1

m¼0ci;2mþð��2b�=2cÞ½q�; ð14Þ

E½ðr½C�i;� ½q�Þ2� ¼1

TS

Z ðqþ1ÞTSþT ½C�i;�

qTSþT ½C�i;�

�2ðtÞdt ¼ �2½C�i;� ½q�: ð15Þ

Then, the receiver de-spreads the sequence of the samples with the spreading code

wm corresponding to the m-th machines, and the interference component disappears

by the orthogonal code. Thus, qi;� correspondence of bi;� is obtained. Based on this

value qi;�½p�, the receiver makes the decision.© IEICE 2019DOI: 10.1587/comex.2019XBL0097Received June 19, 2019Accepted June 28, 2019Publicized July 16, 2019Copyedited October 1, 2019

413


3 BER and PER

The error rate of the p-th bit of the μ-th slot in the i-th period of TDMA is given by

BER½T�i;� ½p� ¼

1ffiffiffiffiffi2�

p�½T�i;� ½p�

Z 1ffiffiffiffi2P

p exp � x2

2�½T�i;� ½p�2 !

dx ¼ 1

2erfc

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP

�½T�i;� ½p�2s !

; ð16Þ

where erfcð�Þ is the complementary error function. Similarly, the error rate of the

p-th bit of the signal of μ-th slot in SCDMA is given by

BER½C�i;� ½p� ¼

1

2erfc

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPM2XM�1

l¼0 �½C�i;� ½p þ lNP�2

vuut0B@

1CA: ð17Þ

The packet error rate PER�½i� becomes

PERi;� ¼ 1 �YNP�1

p¼0ð1 � BERi;�½p�Þ: ð18Þ

4 Numerical examples

In this section, we present numerical examples of BER and PER with system and

noise parameters shown in Tables I(a) and I(b). Since SNR is time-varying, we use

the average SNR (SNR) defined below as the parameter of communication channel

quality

SNR ¼ 1

MTC

ZTC

XM�1m¼0 s

2mðtÞ

E½n2ðtÞ� : ð19Þ

In the transmission of states information, SCDMA has M times transmission time

compared with TDMA. Therefore, under the condition that SNR is the same, the

transmission power of each device of SCDMA is 1=M of TDMA.

Fig. 2(a) and Fig. 2(b) show the BER of each bit within one control period with

TDMA and proposed SCDMA, respectivly. The BER of each bit with TDMA

varies greatly, whereas that with SCDMA is almost equal by chip-level interleav-

ing. Also, since the bit duration of SCDMA is four times longer than that of

TDMA, the influence of the cyclic impulse noise at the 224-th bit of TDMA, which

often dominates the overall system performance, is mitigated as the 61-th bit of

SCDMA.

Table I. Parameters

(a) For communication (b) For the PLC noise (L ¼ 3).

Modulation scheme BPSK ‘ 0 1 2

Average SNR 3∼7 [dB] A‘ 0.230 1.38 7.17

M 4 �‘ [deg] - −6 −35NP 40 [bit] n‘ 0 1.91 1:57 � 105

1=TAC 60 [Hz] a 1:2 � 10�5

Spreading code Walsh

Length of codes 4


414


As shown in Fig. 2(c), the PER in each slot of TDMA varies because of

fluctuation of noise power. Therefore each machine has a different communication

quality. On the other hand, as shown in Fig. 2(d), the PER of all four code channels

of SCDMA are equal. Comparing Figs. 2(c) and 2(d), we find that the slots 3–7 of

TDMA have worse PER than that of SCDMA. This suggests that the machines

using these TDMA slots may have better control performance if SCDMA is used.

5 Conclusion

In this paper, we proposed the introduction of SCDMA with orthogonal codes in

multiple machine control system using PLC, and compare the communication

quality of TDMA. It is confirmed that the proposed system provides a perfectly

equal-quality channel to every machine, while each TDMA channel has quality

difference. In conclusion, SCDMA is suitable for multiple machine communication

because of its perfect equal-channel quality for each machine, ease of synchroniza-

tion using the strong mains voltage as a clock, and good PER performance.

Acknowledgment

The authors would like to express their gratitude to Prof. YAMAZATO of Nagoya

University for his valuable suggestions.

(a) BER (TDMA) (b) BER (SCDMA)

(c) PER (TDMA) (d) PER (SCDMA)

Fig. 2. Calculation result.


415


Neural network basedchannel identification andcompensation

Takaki Omuraa), Shun Kojima, Kazuki Maruta,and Chang-Jun AhnGraduate School of Engineering, Chiba University,

1–33 Yayoi-cho, Inage-ku, Chiba 263–8522, Japan


Abstract: This letter proposes a neural network based channel identifica-

tion and compensation methods for an OFDM system. Under the fast fading

environment, pilot-aided channel estimation suffers from channel state

fluctuation particularly in the last part of the packet. The proposed approach

can estimate the whole transition of channel states and efficiently compensate

the channel variation using the generalization capability of a neural network.

The computer simulation results clarify its effectiveness via improved BER

performance even under the stringent Doppler shift.

Keywords: OFDM, fast fading, nonlinear prediction, artificial neural net-

work, channel estimation

Classification: Wireless Communication Technologies

References

[1] M. Yofune, C. Ahn, T. Kamio, H. Fujisaka, and K. Haeiwa, “Decision direct andlinear prediction based fast fading compensation for TFI-OFDM,” Far East J.Electron. Commun., vol. 3, no. 1, pp. 35–52, July 2007.

[2] S. Soejima, Y. Ida, C. Ahn, T. Omori, and K. Hashimoto, “Fast fadingcompensation based on weighted channel variance for TFI-OFDM,” J. SignalProcess., vol. 17, no. 3, pp. 41–49, May 2013. DOI:10.2299/jsp.17.41

[3] T. Omura, S. Kojima, K. Maruta, and C. Ahn, “Neural network based channelidentification and compensation,” IEEE ISCIT 2018, pp. 349–354, Sept. 2018.DOI:10.1109/ISCIT.2018.8587981

[4] Z. Ghassemlooy, W. Popoola, and S. Rajbhandari, Optical Wireless Communi-cations: System and Channel Modelling with MATLAB, pp. 327–339, Mar.2017.

[5] M. Hagan and M. Menhaj, “Training feed-forward networks with the Marquardtalgorithm,” IEEE Trans. Neural Netw., vol. 5, no. 6, pp. 989–993, Nov. 1994.DOI:10.1109/72.329697

© IEICE 2019DOI: 10.1587/comex.2019XBL0095Received June 11, 2019Accepted July 2, 2019Publicized July 24, 2019Copyedited October 1, 2019

416


http://dx.doi.org/10.2299/jsp.17.41




http://dx.doi.org/10.1109/ISCIT.2018.8587981




http://dx.doi.org/10.1109/72.329697

http://dx.doi.org/10.1109/72.329697

http://dx.doi.org/10.1109/72.329697

1 Introduction

The well known channel estimation method is the pilot-aided channel estimation

(PCE) where known training symbols are inserted in head of the packet. Under

the fast fading environment, significant errors occur due to rapidly changing the

channel state. Estimated channel state information (CSI) is largely dissimilar to the

real channel state particularly in the last part of the packet. In order to mitigate this

impact, a large number of pilot symbols should be inserted frequently to identify

the accurate CSI in compensation for increase of overhead. Accordingly, the

transmission efficiency is degraded and required energy per bit is also increased

because the pilot symbols cannot contribute to information transfer. To overcome

these problems, a data-aided decision feedback channel estimation (DFCE) has

been proposed [1, 2]. By exploiting data part, this method generates the replica of

received signals using the remodulated signals and CSI for given pilot signals.

Channel variation can be compensated by using the error components derived from

the difference between the replica symbols and received symbols. However, this

method requires the accurate remodulated symbols to identify the accurate CSI.

Accordingly, the BER performance is not so improved even if estimated CSIs by

DFCE are applied for a lot of data symbols.

To improve the accuracy of estimated CSI transition, we proposed an artificial

neural network (ANN) approach [3]. The proposed method first applies data-aided

CSI estimation for some data symbols and the trains a neural network. The

generalization capability of a neural network interpolates the whole transition of

channel state and compensates the channel variation. This letter refined overall

evaluation with employing more simplified ANN algorithms than our previous

study [3].

2 Decision feedback channel estimation (DFCE)

Here we assume that SISO-OFDM transmission pilot symbols are inserted at the

head of data symbols. DFCE exploits the demodulated signals and the CSI given

the pilot signals. Let ~HðkÞ and Xðk; iÞ denote the pilot-aided CSI and the i-th

remodulated symbol after the detection/error correction at the k-th subcarrier,

respectively, the replica signal, Yrepðk; iÞ, is generated by

Yrepðk; iÞ ¼ ~HðkÞXðk; iÞ: ð1ÞThe channel variations at the i-th symbol, �Hðk; iÞ, can be calculated by

�Hðk; iÞ ¼ Yðk; iÞYrepðk; iÞ ; ð2Þ

where Yðk; iÞ denotes the original received symbol. The adjusted CSI at the i-th

symbol, ~Hðk; iÞ, is the given by,

�Hðk; iÞ ¼ �Hðk; iÞ � ~HðkÞ: ð3ÞPractically, CSI component includes the noise term which degrades demodulation

accuracy. Here we perform the noise reduction by using adjacent symbols; 3

samples are averaged over �Hðk; i � 1Þ, �Hðk; iÞ and �Hðk; i þ 1Þ;© IEICE 2019DOI: 10.1587/comex.2019XBL0095Received June 11, 2019Accepted July 2, 2019Publicized July 24, 2019Copyedited October 1, 2019

417


Hðk; iÞ ¼Xiþ1

j¼i�1�Hðk; jÞ

3: ð4Þ

In order to trace the whole transition of the CSI, DFCE repeats the CSI estimation

every s symbols. Accordingly, the channel response for the conventional DFCE

method at the i-th symbol, Hdecðk; iÞ, is

Hdecðk; iÞ ¼

~HðkÞ (1 � i � s)

Hðk; s þ 1Þ (s þ 1 � i � 2s)

..

. ...

Hðk:�s þ 1Þ (�s þ 1 � i � Nd)

8>>>>><>>>>>:

; ð5Þ

where β and Nd are the number of subsampling CSI by DFCE and the number of

data symbols, respectively. β can be calculated by dNds e � 1.

However, the above method cannot identify accurate CSI when decision errors

are occurred. Thus, the BER performance is not so improved even if CSI by DFCE

are applied especially under the fast fading environment.

3 Proposed method

The proposed method applies an ANN for channel identification and compensation

using partially obtained CSI via DFCE at constant interval. ANN has a nonlinear

statistical modeling capability for pattern classification and for constructing com-

plex relationships between inputs and outputs extensively [4]. Accordingly, the

generalization capability of the ANN trained by only a few estimated CSI achieves

a high-accuracy tracing whole CSI transition.

Fig. 1(a) represents the block diagram of the proposed channel estimation and

the ANN structure, when the number of subcarriers is K. Ik and Qk (k ¼ 1; 2; . . . ; K )

are applied to k-th subcarrier’s real and imaginary part of the CSI, respectively.

Where e denotes a vector of network errors. The ANN is trained to adjust weights

for regression analysis. The ANN is applied for a fully connected 2-layer feedfor-

ward network. The multilayer feedforward neural network is one of the represen-

tative structures of neural networks.

Fig. 1(b) presents the process of the proposed method. Before the ANN

training, CSIs are estimated every s symbols: the first one is by PCE and remaining

are by DFCE. We can set these information as the desired responses and

½1; s þ 1; . . . ; �s þ 1� as the training input.

In this letter, the Levenberg-Marquardt algorithm is applied for the ANN

training. This algorithm was designed to approach second-order training speed

without computing the Hessian matrix. The algorithm appears to be the fastest

method for training moderate-sized feedforward neural networks (up to several

hundred weights).

The algorithm is like a combination algorithm of steepest descent and the

Gauss-Newton method. The Gauss-Newton method may happen an oscillation

and a divergence when the current solution is far from the global minimum. The

Levenberg-Marquardt algorithm is one of the methods for handling divergence of

the Gauss-Newton method. When the current solution is far from the optimum


418


solution, the algorithm behaves like a steepest descent method: slow, but guaran-

teed to decrease a performance function at each iteration. When the current solution

is close to the optimum solution, it becomes like a Gauss-Newton method. Further

details of this algorithm can be found in [5].

After the ANN training, all data symbols indice are serially input to the

ANN. Consequently, we can interpolate the whole transition of CSI, Hout ¼½houtð1Þ;houtð2Þ; . . . ; houtðNdÞ�, and use these informations for the channel compen-

sation. Where houtðiÞ (i ¼ 1; 2; . . . ; Nd) denotes the trained ANN based all sub-

carriers’ estimated CSI applied for i-th data symbol.

4 Computer simulation results

Table I presents the simulation parameters. Where Np denotes the number of pilot

symbols. We employed the Jakes’ model to represent time varying channel where

direction of arrival of 16 incoming rays are uniformly distributed. The maximum

Doppler frequency is 700Hz. Its normalized value is 2:8 � 10�3.First, we examine the best parameters for our proposal in this simulation

condition. Figs. 2(a), 2(b), and 2(c) show the BER performances versus each

parameters of the DFCE and proposed method (Prop.) at Eb=No ¼ 25 dB. Ne, Mg,

and Nn denote the maximum number of epochs for the ANN training, the minimum

(a) Block diagram of the proposed channel estimation (1:ANN training 2:re-gression analysis) and architecture of the ANN

(b) The process of the proposed method (1:before 2:after) ANN training

Fig. 1. Proposed method


419


performance gradient for the ANN training, and the number of neurons for the

ANN hidden layer, respectively. From these performance comparisons, we employ

parameters as s ¼ 7 for DFCE and s ¼ 12, Ne ¼ 10,Mg ¼ 10�7, and Nn ¼ 1 for the

proposed method to achieve the best performances. In this case, β is calculated as

2 for DFCE and 1 for the proposed method. These optimized results imply that the

generalization capability of ANN requires only two CSI estimate values to yield the

best BER performance.

Fig. 2(d) shows the BER performances of the conventional and the proposed

methods, respectively. The BER of the PCE and the DFCE shows the error floor.

The PCE cannot track the whole transition of CSI and even the DFCE cannot

compensate it; decision errors are significant. On the other hand, the proposed

method can remove the error floor and satisfactorily achieve BER below 10�4. Thisis because the generalization capability of the ANN can compensate such a drastic

channel variation. Accordingly, our proposed ANN based channel compensation is

quite effective in terms of channel tracking performance, even in high mobility

environment.

Table I. Simulation parameters

Transmission scheme OFDM

Data modulation QPSK

IFFT size, Number of carriers 64

Guard interval 16

Number of pilot/data symbols Np=Nd ¼ 2=20

Fading 15 path Rayleigh fading, 1 dB decay

Max Doppler frequency 700Hz

Transmission bandwidth 20MHz

Forward error correctionConvolutional code(R ¼ 1=2, K ¼ 7)

Activation function (hidden layer) Logistic sigmoid

Activation function (output layer) Linear

Learning algorithm Levenberg-Marquardt backpropagation


420


5 Conclusion

This letter proposed the channel estimation and compensation method using ANN

to track the whole transition of CSI under the fast fading environment. Due to the

generalization capability of the ANN, the proposed method has shown improved

BER performance compared to the conventional PCE and DFCE even at high

mobility environment as Doppler frequency of 700Hz.

(a) BER vs. adaptive interval s for DFCEand the proposed method (Eb/No = 25 dB)

(b) BER vs. maximum number of epochsNe and number of neurons Nn for the pro-posed method (Eb/No = 25 dB)

(c) BER vs. minimum performance gradi-ent Mg for the proposed method (Eb/No =25 dB)

(d) BER vs. Eb/No

Fig. 2. Simulation results


421


Detection schemes formassive MIMO system withlow-resolution ADCs

Peng Gaoa) and Yukitoshi Sanadab)

Dept. of Electronics and Electrical Engineering, Keio University,

3–14–1 Hiyoshi, Kohoku, Yokohama 223–8522, Japan


b) [email protected]

Abstract: In a full-digital massive multi-user MIMO system, maximal-ratio

combining (MRC) can obtain more considerable diversity gain while inter-

stream interference (ISI) and multi-user interference (MUI) can be canceled

using minimum mean square error (MMSE) algorithm. This letter evaluates

the throughputs of detection schemes for different antenna numbers in the

massive MIMO system with low-resolution analog-to-digital converters

(ADCs). The letter makes a comparison between MRC and MMSE under

quantization range limit. Numerical results show that MRC achieves better

system performance with lower implementation complexity as the number of

antennas increases.

Keywords: massive MIMO, low-resolution ADCs, MRC, MMSE, range

limit

Classification: Transmission Systems and Transmission Equipment for

Communications

References

[1] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, andJ. C. Zhang, “What will 5G be,” IEEE J. Sel. Areas Commun., vol. 32, no. 6,pp. 1065–1082, June 2014. DOI:10.1109/JSAC.2014.2328098

[2] D. Dardari, “Joint clip and quantization effects characterization in OFDMreceivers,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 53, no. 8, pp. 1741–1748, Aug. 2006. DOI:10.1109/TCSI.2006.875170

[3] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, and C. Studer, “Throughputanalysis of massive MIMO uplink with low-resolution ADCs,” IEEE Trans.Wireless Commun., vol. 16, no. 6, pp. 4038–4051, June 2017. DOI:10.1109/TWC.2017.2691318

[4] A. Azizzadeh, R. Mohammadkhani, and S. V. A. Makki, “BER performance ofuplink massive MIMO with low-resolution ADCs,” IEEE ICCKE, Oct. 2017.DOI:10.1109/ICCKE.2017.8167895

[5] T. E. Bogale and L. B. Le, “Beamforming for multiuser massive MIMO systems:Digital versus hybrid analog-digital,” IEEE Global Communications Confer-ence, Dec. 2014. DOI:10.1109/GLOCOM.2014.7037444

[6] S. Yoshioka, S. Suyama, T. Okuyama, J. Mashino, and Y. Okumura, “5G massiveMIMO with digital beamforming and two-stage channel estimation for low SHFband,” IEEE Wireless Days, Mar. 2017. DOI:10.1109/WD.2017.7918124

© IEICE 2019DOI: 10.1587/comex.2019XBL0105Received July 19, 2019Accepted July 29, 2019Publicized August 8, 2019Copyedited October 1, 2019

422






http://dx.doi.org/10.1109/TCSI.2006.875170




http://dx.doi.org/10.1109/TWC.2017.2691318





http://dx.doi.org/10.1109/ICCKE.2017.8167895




http://dx.doi.org/10.1109/GLOCOM.2014.7037444




http://dx.doi.org/10.1109/WD.2017.7918124

http://dx.doi.org/10.1109/WD.2017.7918124

http://dx.doi.org/10.1109/WD.2017.7918124

http://dx.doi.org/10.1109/WD.2017.7918124

1 Introduction

In the next generation mobile communication system, much higher area data

throughput is required to manage the global demand for the continuously growing

wireless data traffic [1]. It can achieve a multiple times larger bit rate by improving

spectral efficiency (bits/s/Hz/cell) without the need for more bandwidth or addi-

tional base stations. Since a large number of the base station (BS) antennas

effectively averages out noise and fading, and reduces the multi-user interference

to a certain extent, massive MIMO is considered as a key technology.

Full digital massive MIMO deployment is difficult to realize because high-

resolution analog-to-digital converters (ADCs) produce primary power consump-

tion. Owing to its favorable property of low cost and low power consumption, low-

resolution ADCs (1∼4 bits) have also been worth paying attention [2, 3, 4].

Reference [2] proves that an appropriate quantization range limit (clipping) can

relieve the distortion in a single-input single-output OFDM system. In [3], the

performance of single-carrier modulation in a massive MIMO system with low-

resolution ADCs is analyzed while OFDM is assumed in this letter. In [4], it is

shown that a minimum mean square error (MMSE) algorithm can achieve better

BER using low-resolution ADCs in a hybrid analog-digital system with a large

number of BS antennas. MMSE detection demands a large amount of computa-

tional complexity for matrix inversion while it achieves near-optimal performance.

Maximal-ratio combing (MRC), as one beamforming scheme, can also be applied

at the receiver of full-digital massive MIMO to achieve the maximum receive

signal-to-interference-plus-noise ratio (SINR) [5].

In this letter, the achievable uplink rates of MRC and MMSE with the

quantization range limit and with the different numbers of BS antennas are

evaluated and compared. Residual multi-user interference (MUI) caused by low-

resolution ADCs limits the system throughput even though MMSE is applied. On

the other hand, when low-resolution ADCs are used, MRC achieves comparable

throughput performance as that of MMSE. Thus, MRC is more suitable owing to its

lower complexity.

2 System model of uplink

2.1 Uplink system model

The single-cell uplink system shown in Fig. 1 is assumed. We consider a full-

digital massive MIMO uplink with low-resolution ADCs. There are K users, each

Fig. 1. Uplink system of massive MIMO with low-resolution ADCs.


423


user has N Tx antennas and one BS equipped with an array of M antennas. Suppose

that the size of an inverse discrete Fourier transform (IDFT) is Ndft, and the signal

from the nth antenna of the kth user in the uplink is,

skn½u� ¼ 1ffiffiffiffiffiffiffiffiNdft

p XNdft�1v¼�Ncp

Skn½v�ej2�uv=Ndft ; ð1Þ

where Skn½v� is the transmit signal from the nth antenna of the kth user on the vth

subcarrier, Ncp is the length of a cyclic prefix, and skn½u� is the transmit signal at the

uth time index. The received signal of the mth antenna of the BS is given as

ym½u� ≜XK

k¼1XN

n¼1ðffiffiffiffiffiPk

phmknskn½u�Þ þ zm½u�; ð2Þ

where the power of skn½u� is E½jffiffiffiffiffiPk

pskn½u�j2� ¼ 1, Pk is the transmit power of the

kth user, hmkn is the channel response between the nth antenna of the kth user and

the mth antenna of the BS, and zm½u� � CNð0; �2ULÞ is the thermal noise.

With the assumption of identical low-resolution ADCs, the in-phase and

quadrature components of the received signal of the mth antenna are quantized

by the ADCs of b-bit resolution as follows,

qm½u� ≜ QðRðym½u�ÞÞ þ jQðImðym½u�ÞÞ; ð3Þwhere Qð:Þ represents quantization; Rð:Þ and Jð:Þ denote the real and imaginary

parts, respectively. The quantization through the ADCs worsens the accuracy of

channel estimation and deteriorates the system performance. The quantization noise

is given as,

em½u� ¼ qm½u� � ym½u�: ð4ÞThe receiver removes the cyclic prefix and put into a discrete Fourier transform

(DFT) block. The signal on the vth subcarrier is then given as,

Ym½v� ¼XK

k¼1XN

n¼1ffiffiffiffiffiPk

pHmkn½v�Skn½v� þ ðZm½v� þ Em½v�Þ; ð5Þ

where,

Skn½v� ¼ 1ffiffiffiffiffiffiffiffiNdft

p XNdft�1u¼0 skn½u�e�j2�uv=Ndft ; ð6Þ

Hmkn½v� ¼ hmkne�j2�v=Ndft ; ð7Þ

Em½v� ¼ 1ffiffiffiffiffiffiffiffiNdft

p XNdft�1u¼0 em½u�e�j2�uv=Ndft ; ð8Þ

and Zm½v� � CNð0; �2ULÞ is the thermal noise that is given by,

Zm½v� ¼ 1ffiffiffiffiffiffiffiffiNdft

p XNdft�1u¼0 zm½u�e�j2�uv=Ndft : ð9Þ

2.2 Channel estimation

Least-square (LS) estimation is employed for channel estimation [6] as follows.

Suppose that data symbols are mapped at the subcarriers, which are the ð�Þth to the

ð� þ Nsc � 1Þth subcarrier, where Nsc is the number of active subcarriers, and a

channel response is estimated for each subcarrier. The transmit active subcarriers

are divided as B ¼ ðNsc=V Þ blocks in frequency domain, where V is the number of

subcarriers in one block. Nsym ¼ L=V OFDM symbols are required when the

uplink pilots for L streams are inserted at each V subcarriers to estimate channel

responses.


424


The channel in the bth block is estimated during a channel estimation period.

The subcarrier index, fvg, in the bth block is from ðb � 1ÞV þ � to bV . The

estimated channel for the bth block is

Hb

mkn ¼1

V

XbV�1þ�v¼ðb�1ÞVþ�ðSkn½v�Þ

�1Ym½v�; ð10Þwhere Skn½v� and Ym½v� are the orthogonal sequence and the received signal on the

vth subcarrier, respectively.

2.3 Uplink performance

In the uplink, an ergodic rate per arbitrary user k is

CULk ¼

XN

n¼11

BV

XB

b¼1XbV�1þ�

v¼ðb�1ÞVþ� log2ð1 þ SINRvknÞ; ð11Þ

where

SINRvkn ¼

jEfWbknHkn½v�gj2XN

n¼1XK

i¼1i≠k

EfjWbknHin½v�j2g þ

XNj¼1j≠n

jEfWbknHkj½v�gj2 þ �0UL

2; ð12Þ

�0UL2 ¼ EfkWb

knk2g�2UL þ jEfWbknEm½v�gj2; ð13Þ

Wbkn is the maximum ratio combining or minimum mean square error detection

coefficients in the vector form, they are given as

MRC:

Wbkn ¼ ðHb

knÞH ¼Hb

1kn

..

.

HbMkn

264

375

0B@

1CA

H

; ð14Þ

Or MMSE:

Wbkn ¼ ðððHbÞðHbÞH þ �0UL

2IÞ�1ðHbknÞÞH; ð15Þ

where Hb and Hbkn are the channel estimation response including estimation error, in

the b-th block and Hkn½v� is the channel responses between the nth antenna of the

kth user and the base station antennas in the vector form,

Hij½v� ¼H1ij½v�

..

.

HMij½v�

264

375; ð16Þ

E½v� is the quantization noise vector given as

E½v� ¼E1½v�...

EM½v�

24

35: ð17Þ

Finally, �0UL2 is the sum of the variance of the thermal noise and the quantization

noise.

2.4 Proposed ADC design

In the proposed ADC, the quantization range is given as

Amax ¼ C � Eðjym½u�jÞ; ð18Þ

� ¼ 2Amax

2R; ð19Þ

where C is the coefficient of the quantization range that is relative to Eðjym½u�jÞ,jym½u�j is derived through the output of the power detector equipped in each


425


antenna of the BS, and Δ is the quantization step-size. The quantization range is

adjusted via the coefficient C and the quantization converts the real input signal to a

real-valued output, ri, for i ¼ 1; 2; . . . ; 2R. The ith output value after ADCs is

defined as

qi ¼ � 2R

2� 1

2þ i

� ��: ð20Þ

Thus, its value is adjusted for suppressing extra noise. Since each antenna element

of the massive MIMO system generally receives a weak signal, the amplitude of

which is smaller than that of the thermal noise and limiting the quantization range

improves the signal quality after combining in the receiver.

3 Numerical results

3.1 Simulation conditions

Computer simulation conditions are presented in Table I. The massive MIMO BS

receives signals with M ¼ 128 or 1024 antenna elements. In MU transmission,

there are eight users (each user with Nu ¼ 2 antennas) that communicate with the

BS simultaneously. The antenna spacing is 0:5� at the BS and 1:0� at each user,

where λ is the wavelength. The number of active subcarriers is 1200 while the DFT

size is 2048. The number of blocks is 150 and the number of subcarriers per block

is 8. Since the number of symbols for channel estimation is two, the number of

signal streams whose channel responses can be estimated with Zad-off Chu

sequences as orthogonal sequences is 16. As a channel model, i.i.d. Rayleigh

fading is assumed. The resolution of ADCs is selected from 1, 2, 3, 4, or infinite

bits. System throughput is the total rate of eight users. The number of trials for each

plot is 10000.

3.2 System performance

Performance of MRC and MMSE with low-resolution ADCs under different

amounts of antennas is presented in Fig. 2. As shown in Figs. 2(a) and (b), when

Table I. Simulation conditions

Bandwidth 20MHz/RB

Number of Antennas Massive MIMO: M ¼ 128; 1024(¼ horizontal 8,64 � vertical 16),

Each user: N ¼ 2(¼ horizontal 1 � vertical 2),

Number of Users K ¼ 8

Interval of antennas BS: 0:5�, UE: 1�

Number of Subcarriers nsc ¼ 1200 (DFT Points: Ndft ¼ 2048)

Number of Blocks B ¼ 150

Number of Subcarrier/Block V ¼ 8

Number of Symbols in Pilot Slot Np ¼ 2

Pilot Sequence Zad-off chu sequence (Length 16)

Number of symbols for Channel Estimation 2

Channel Model i.i.d Rayleigh Fading

ADCs’ resolution 1,2,3,4 and infinite

Number of Trials 10000 channel responses/plot


426


the number of antennas is 128, the performance with infinite ADCs is not only

limited by MUI, but also the over-clipping of the original signal deteriorates the

channel estimation accuracy. With infinite ADCs, MMSE detection is better due to

MUI elimination.

Because of a limited number of BS antennas, low-resolution ADCs has

produced severe bit error rate (BER) in a TDD OFDM system. In a massive

MIMO system, a large number of BS antennas can be equipped to reduce the BER.

However, the MMSE signal detection algorithm involves matrix inversion with

large complexity (especially in massive MIMO). MRC can avoid complicated

matrix inversion. In Fig. 2(c) and (d), they are clear that the performance of MRC

with 1024 antennas is as good as MMSE. MRC can achieve even better perform-

ance than that of MMSE with 1, 2, 3-bit ADCs.

4 Conclusions

The different signal detection algorithms used by the uplink receiver in a multi-user

massive MIMO system are compared. Under a large number of BS antennas,

MMSE with higher resolution ADCs is better than MRC, although the matrix

inversion operation demands a large amount of computational complexity. On the

other hand, MRC can achieve equivalent or even better performance, especially

with low-resolution ADCs. The system throughput realized in MRC with 2-bit

ADCs is almost the same as that with infinite resolution ADCs and MRC can

achieve better performance than MMSE with 1, 2, 3-bit ADCs. In addition, it

requires less computational complexity. Therefore, MRC is more suitable as the

number of BS antenna elements increases in a full-digital massive MIMO with

low-resolution ADCs.

(a) MRC

(c) MRC

(b) MMSE

(d) MMSE

Fig. 2. Performance of MRC and MMSE with low-resolution ADCs


427


ieice communications express, vol.8, no.10, 404 comparison ... · is a typical usage scenario in...

Documents