Full-duplex prototype based on jointpassive and digitalcancellation method

DiWu,Can Zhang,*,Shaoshuai Gao,and Heping Zhao

1.Schoolof Electronic,Electricaland Communication Engineering,University of Chinese Academy of Sciences, Beijing 100049,China;

2.China Academy of Space Technology,China Aerospace Science and Technology Corporation,Beijing 100081,China

Full-duplex prototype based on jointpassive and digitalcancellation method

DiWu1,Can Zhang1,*,Shaoshuai Gao1,and Heping Zhao2

1.Schoolof Electronic,Electricaland Communication Engineering,University of Chinese Academy of Sciences, Beijing 100049,China;

2.China Academy of Space Technology,China Aerospace Science and Technology Corporation,Beijing 100081,China

Recentresearch shows thatitis possible to achieve the full-duplex system by cancelling strong self-interference signals, which can be divided into three classes respectively,i.e.,passive cancellation,active cancellation and digital cancellation.This paper tries to achieve the full-duplex system without using active cancellation,thus a full-duplex system using a joint mechanism based on a novelpassive cancellation method and a noveldigital cancellation method is proposed.Therein,a good antenna placementguided by the theory of the antenna electromagnetic field for the passive cancellation is presented.For the proposed digitalcancellation method,unlike previous separate mechanisms,it is designed by using the recursive least square(RLS)algorithm jointly with passive cancellation.The self-interference channel state information(CSI)is transferred as the inputof digitalcancellation to balance the performance and the complexity.Experimentalresults show that the proposed self-interference cancellation mechanism can achieve about 85 dB which is better than the previous research.Meanwhile,this design provides a better performance compared with half-duplex with both line-of-sightchanneland nonline-of-sightchannel.

full-duplex,self-interference cancellation,jointpassive and digitalmechanism.

1.Introduction

As wireless communication develops rapidly,everincreasing demands on the limited wireless spectrum is needed which drives the quest for systems with a higher spectralefficiency.Currentdeployed wireless communication systems,such as time-division duplex system or frequency division duplex system,cannotoperate in the fullduplex mode,i.e.,they cannottransmitand receive simultaneously in the same frequency band.The reason is that the signalfrom a localtransmitting antenna which is called the self-interference signalis much strongerthan the signal ofinterestfrom other nodes.As a resultof the large power differential,the signalis swamped by the self-interference. If the self-interference can be removed,the node can work in the full-duplex mode which will offer the potential to double the spectral efficiency.Consequently,the problem has attracted extensive attention from both industry and academia,and has caused significantfollow up work.

To achieve full-duplex,the significant self-interference of a node must be cancelled completely.Recent results [1–11]have demonstrated the feasibility of full-duplex wireless communication using different kinds of separate mechanisms to reduce the self-interference since 2010.In general,they can be classified into three classes in three domains respectively,i.e.,passive cancellation in the propagation domain,active cancellation in the analog circuit domain and digitalcancellation in the digitaldomain.

The aim ofpassive cancellation is to isolate the transmitting chain from the receiving chain electromagnetically.It accounts for a large portion of the total self-interference suppression in existing full-duplex designs.Choi etal.[1] and Aryafar et al.[9]placed a single receiving antenna at a precise location where the carrier waveforms were antiphase.Itis notidealforbroadband signals.Ifthe receiving antenna is placed perfectly forthe centerfrequency,there is a smallerror in placementfor the other frequencies within bandwidth.By building a system with vertically polarized receiving antenna and horizontally polarized transmitting antenna,Aryafaretal.[9]and Everettetal.[11,12]demonstrated thatcross-polarization offered an additionalmechanism to electromagnetically isolate the transmitting and receiving antennas.However,in indoor environments,as signals bounce off walls and other objects,polarity of signals willchange which willresultin a worse performance [12].Meanwhile,the directionalantennas were adopted in[3,11,12]which can isolate transmitenergy away from the receiving antenna.However,six directional antennas are needed to achieve 360°pattern which is too complex[11].

The aim of active cancellation is to suppress selfinterference in the analog receiving chain before the analog-to-digital converter(ADC).According to the research from Rice University,the additionalhardware components required for active analog cancellation consist of one digital-to-analog converter(DAC),one transmitter radio and one radio frequency(RF)adder.Likewise,as an additional hardware,the QHx220 was used in the fullduplex system[1,2,10].In QHx220,the amplitude and phase of the interference reference signal can be adjusted adaptively to match the interference signalcontained in the received mixed signal.However,the hardware constraints in QHx220 lead to non-linear distortion[1].Similarly,by using extra hardware components,Bharadia etal.designed a 10 cm×10 cm printed circuit board(PCB)to conduct active cancellation[8].

Digitalcancellation aims to remove the self-interference after the ADC in the baseband.Its cancellation performance is decided by the accuracy of the channel estimation.Previous work[1,13]can achieve about 10–15 dB. The researchers from Stanford[2]applied digital cancellation by using the least squares algorithm to estimate the channel frequency response which can reduce selfinterference by up to 30 dB with a transmit power of 0 dBm.However,its performance declines by up to 9 dB athigherreceived powers.

Although the active cancellation techniques are attractive due to its extra cancellation,the effectivenessis greatly limited by the device size,i.e.,the smaller the device,the less room to implementsuch PCB or analog chain and RF adders.It should be noted thatthere are some options and trade-offs in the design of active cancellation techniques. Each method above introduced additional hardware devices which introduce extra non-liner distortion and noise. Meanwhile,according to[13],it has been found that as the performance of active analog cancellation gets better, the effectiveness of digital cancellation after active analog cancellation reduces,sometimes this even increases the self-interference.

However,compared with active cancellation,passive and digital cancellation methods have their superiorities. Passive cancellation accounts for a large portion of the totalself-interference suppression in existing full-duplex designs,and its methods are simple to achieve.As for suppression methods in the digitaldomain,itis relatively easy to conductdigitalsignalprocessing with the rapid developmentofmicroelectronics industry.Nevertheless,according to the previous research[1,2,4],passive and digitalcancellation methods are designed separately from other cancellation methods,which means the digitalmethod could not adapt itself based on others’performance.From the discussion above,we try to achieve full-duplex radio using a jointmechanism based on passive and digitalcancellation methods withoutthe active cancellation method.

The first contribution of this paper is that we design a joint mechanism based on a new passive cancellation method and a new digital cancellation method.For the passive method,we find out where the power null point of dipole antenna is,and propose a simple but effective mechanism.As[12]showed that passive suppression would cause a decrease in the coherence bandwidth of the residual self-interference channel.As a result,we design a self-interference channel state information(CSI)transfer mechanism to join the two cancellation methods as Fig.1 shows.The change of self-interference CSI caused by the passive method will be the input of the proposed digital cancellation algorithm.The algorithm will adjust adaptively to balance the cancellation effect and computationalcomplexity.

Fig.1 Joint passive cancellation and digitalcancellation method

The second contribution of this paper is thatwe present a real-time prototype full-duplex infrastructure node which is implemented in the wireless open-access research platform lab(WARPLab)[14]framework.Our experiments demonstrate that the proposed design is low costand easy to build and can achieve 85 dB of cancellation on average. The cancellation result is even better than prior separate designs using three steps.Further,this prototype is used to conduct real-time full-duplex experiments.The results show thatits performance outperforms the half-duplex system noticeably.

The organization of the rest paper is structured as follows.The jointcancellation mechanism design is described in Section 2.In Section 3,the prototype implementation and experiment scheme based on WARP boards is presented.Evaluation ofourdesign and comparisons with previous researches and half-duplex system are given in Section 4.We conclude this paper with some finaldiscussions in Section 5.

2.Jointcancellation mechanism design

2.1 Passive cancellation method

Passive cancellation is achieved by maximizing the attenuation of the self-interference signal.Six directional antennas were used in[3,11,12]to ensure that the access point(AP)has 360°pattern which costed much compared with the single omni antenna.Meanwhile,in prior studies[9,11,12],cross-polarization was introduced to cancel the interference.However,polarization of the wave is changed by reflection.As a result,despite perfectly crosspolarized transmitting and receiving antennas,the reflected self-interference may not be cross polarized to the receiving antenna.Meanwhile,according to[15],considering the ground effects,the received signalpower willbe different forverticaldipole antenna and horizonaldipole antenna.If the emission is polarized horizontally,there is always zero received power on the horizontal plane.As public users are near ground,horizontal polarized emissions would be poorly received.This is the reason why almost all public address radio emissions have verticalpolarization.In conclusion,we selectomnivertical polarization dipole antennas as our transmitting and receiving antenna.However, how to place them to achieve better performance?

For the dipole antenna,the radiation intensity can be written as follows[15]:

where I0is a constants the permeability of free-space and ε0is the permittivity of free-space space,l is the diploe length,andθis the sphericalcoordinate.

The normalized elevation power patterns with different l are shown in Fig.2.Itis clearthatthe areas above and under the dipole antenna are deep fading areas where the signalis almostimpossible to be received.This phenomenon guide us to find a new passive cancellation method,thatis, in one terminalnode,the transmitting antenna places vertically above the receiving antenna and in another node,the transmitting antenna places vertically under the receiving antenna.And each transmitting-receiving antenna pair is placed in a horizontalplane.In theory,this mechanism can guarantee thatthe receiving antenna can receive the signal of interest as much as possible while the self-interference signal will be received as less as possible.Furthermore, from the Table 1 and Fig.2,we can see thatas the length of the antenna increases,the beam becomes narrower and the directivity increases.However,as Fig.2 shows,when the length of the dipole increases beyond one wavelength (i.e.,l＞λ),the number of lobes begin to increase,which means there willbe signalleakage in the verticaldirection. Thatis,when the length of the dipole l=λ,the elevation power can be minimized in the vertical direction,which means that l=λis the best selection.In practice,considering the whole antenna length,if we take this passive cancellation method to achieve full-duplex,we should select appropriate dipole length based on different frequencies.Forexample,if we select l=λ=12.5 cm to conduct full-duplex on 2.4 G band,the whole length of the antenna is too long and inconvenient.Thus,for2.4 G band,a better tradeoff is l=λ/2=6.25 cm,butfor 5 G band,the best choice is l=λ=6 cm.

Fig.2 Elevation plane amplitude patterns for a thin dipole with different l

Table 1 3 dB beamwidth with different l

2.2 Self-interference CSI transfer

According to previous researches[1,2,4],the digital method is designed independent.As we know,the joint mechanism has shown its superiority in the encoding/decoding domain[16,17].Inspired by this,we propose a joint mechanism in the full-duplex communication system.The parameter of the proposed digital method can change adaptively based on self-interference CSI.

The passive suppression will transform a relatively frequency-flat self-interference channel to a highly frequency-selective channel[12].In order to verify the effectiveness of this conclusion on our passive method,the spectrum analyzer is used to observe the received signal poweroverdifferentfrequencies.Fig.3 is the received signalpower spectrum snapshotatthe receiving antenna(the same node)fortwo cases-the received power withoutany passive cancellation and the received power with our passive cancellation.

Fig.3 Spectrum snapshots showing the effectofpassive cancellation

From Fig.3,the self-interference channel is changed from a flat fading channel to a frequency selective fading channel.This result is consistent with the conclusion in [12].The metric mostoften used to quantify the degree of frequency-selectivity ofa wireless channelis the rootmean square(RMS)delay spreadστand coherence bandwidth Bc.And they are difficult to be measured directly.However,they are easily to be computed based on the power delay profile of the channel P(τ)which can be measured directly based on pilots,whereτrepresents the channel multi-path delay.In[18],the RMS delay can be computed

These delays are measured relative to the firstdetectable signal arriving at the receiver atτ0=0.The coherence bandwidth Bcis a defined relation derived from the RMS delay spread.Coherence bandwidth is a statisticalmeasure of the range of frequencies over which the channelcan be considered“flat”(i.e.,a channelwhich passes all spectral components with an approximately equalgain and a linear phase).Then the relationship between the coherence bandwidth andστis given as follows:

And the relationship between the filter length M andis positive correlation,where BSIis the bandwidth of the self-interference signal(i.e.,transmitted signal),??is the ceiling function.

It should be noted that there is no specific definition for Bcand M.The frequency correlation function and filter length should be selected based on different cases.

Our experimental results show that1 is a relaxed and appropriate value in our digitalcancellation and environment,i.e.,the coherence bandwidth is defined as the bandwidth overwhich the frequency correlation function is above 0.5.

From we mentioned above,the multi-path parameters (i.e.,στ,Bc,M)of self-interference channel are used to the join passive and digital method.The experimental results show that the CSI information transfer can balance the performance of digital cancellation and computation complexity well.

2.3 Digitalcancellation method

The basic idea of digital cancellation is that the selfinterference channeland hardware effects can be modelled as a linear function of the transmitted signal T which is known in the digital domain.

the estimated sample of the transmitted sample,w be the M×1 estimated coefficients vector w=[w(1),w(2),...,w(M)]T,where M is the filter length,u=[T(n),T(n?1),...,T(n?M+1)]Tbe the inputvectorforthe recursive-least-squares (RLS)algorithm to calculate?T(n).Thus,the equation can be written as follows:

The method consists of two steps,i.e.,self-interference channelcoefficients w estimation and digital cancellation implementation.

2.3.1 Self-interference channelcoefficients estimation In our work,the adaptive RLS algorithm is introduced to estimate the self-interference channel.The coefficientvector w of the RLS can continually adjust on a step-by-step basis during the estimation using the pilots.The algorithm to estimate channelcoefficients is shown in Algorithm 1.

Algorithm 1Coefficient vector w computation based on adaptive RLS algorithm

1://Input:forgetting factorλ,filter length M,transmitted pilots Tpil,received pilots Rpil,initial valueδis a smallconstant,w(0)=[0,0,...,0]; 2://Output:coefficientmatrix W;

3:Update the inputvector based on Tpil,u(n);

4:Compute the Kalman gain vector k(n);

5:Compute the filter output using the previous set of filter coefficients w(n?1);

6:Compute the error err(n);

7:Update the filter coefficients forthe nextiteration;

8:Save the new w(n)in the matrix W;

9:Update the matrix P?1for the nextiteration;

2.3.2 Digitalcancellation implementation

In theory,the coefficients of the last iteration should be around the optimal choice(i.e.,may not be the best)on the premise of convergence.Meanwhile,different from the simulation,real-time wireless channel types are timevarying channels.As a result,the forgetting factorλis used to get a better final coefficient wfin our method.In the RLS algorithm,λdefines the system memory and effects the convergence and the ability of the filter to track timevarying statistics in the inputsequence.In our design,wfis defined as follows:

where p is the end index of convergence and q is the start index of convergence.The equation weights coefficient which is close to the payload heavier.Thus,the estimated transmitted samplesAfterthat,the terminal node subtracts the estimated transmitted signal from the received samples.

3.Implementation

As Fig.4 shows,the experiments were conducted in a nearly vacant room.The retort stand,placed on a 70 cm high desktop,is used to fix the two antennas ofeach nodes. Both line-of-sight channel and non-line-of-sight channel can be created with our setup.Considering the length of the antenna and the antenna theory we described above,we select the omni vertical polarization dipole antenna with dipole length l=λ/2=6.25 cm.The whole length of the antenna is 19 cm.

Fig.4 Real-time experiment setup of the full-duplex mechanism

Fig.5 shows ourfull-duplex experimentalsystem.PAis power amplifier and SMA is subminiature version A connector.As the figure shows,our full duplex infrastructure node was prototyped using the WARPLab framework[14]. In the WARPLab framework,baseband signals generate and received signals process(i.e.,digital cancellation,demodulation)in Matlab,but RF waveforms are transmitted over-the-air in real-time using the WARP platform.This framework facilitates experiment implementation by allowing the use of Matlab for the digital signal processing.Allexperiments were conducted ata 2.472 GHz Wi-Fi channelwithoutany other concurrenttraffic.

Fig.5 Diagram of our full-duplex experimental system

Each packetconsists of 14 136 samples which contains 1 024 pilots at the start of the packet and another 1 024 pilots atthe end ofthe packet.Convolutionalchannelcode with coding rate 2/3 is used to correct bit errors just like [5].In our experiments,the transmission power of the WARP radios is characterized by connecting the radios directly to a Agilentsignalanalyzer.

4.Evaluation

4.1 Cancellation and real-time

We evaluate the cancellation performance by measuring how much itcan attenuate the self-interference signal.

4.1.1 Passive cancellation

We first use the spectrum analyzer to calibrate and set the transmit power of different signals at 15 dBm.Then,we connectdevices as shown in Fig.6 and adjustthe position of antennas(i.e.,d)to measure the performance.To compare differentbandwidth signalcancellation performances, we use two signals,i.e.,5 MHz quadrature phase shift keying(QPSK)signal and 20 MHz orthogonal frequency division multiplexing(OFDM)signal.

Fig.6 Passive cancellation modeland block ofmeasurement

Fig.7 shows that the received signal power versus distance d as shown in Fig.6.From the figure,we can see that the received power over 0＜d＜5 cm is obviously higher than that over d＞5 cm.The reason is the power leakage from the lobes.The 3 dB beamwidth of l=λ/2 is 78°,which means the directivity is notgood and the radio pattern is not quite flat.Thus,if the two antennas are too close,the performance will deteriorate rapidly.

As for d＞5 cm,we think there are two reasons contributed to the passive cancellation performance,that is,antenna placement and multi-path effect.We take 5 MHz QPSK signal as an example to illustrate.Line A (–14.9 dBm)stands for the received QPSK signal power strength without any cancellation skill.Line B (–36.1 dBm)stands for the mean received QPSK signal power over d＞5 cm using our antenna placement. Line C(–44.12 dBm)stands for the mean received QPSK signal power over deep fading region(15 cm＜d＜20 cm)which means the passive cancellation performance can achieve.The first part between Line A and Line B means antenna placementcan bring 21 dB extra gain.The second part between Line B and Line C means the multipath effectcan bring 8 dB extra gain.

Fig.7 Received signalpower ofdifferent bandwidth signal

The reason for the second partis the random phase and amplitudes of the different multi-path components cause fluctuations in signal strength.Let aiandθi(t,τ)be the real amplitudes and phase,respectively,of the i th multipath component at the time t.Normally,the line-of-sight (LOS)dominating component a0is much larger than ai(i＞0)which means that the multi-path effect is not obvious.However,the proposed passive cancellation method makes a0smaller a lotwhich means itcould notbe a dominated factor.Then,the multi-path signalcan affect the received power easily.As the receiving antenna is moved over a local area,the self-interference channel changes, and the received signal power will vary governed by the fluctuations of aiandθi.Therein,aivaries little over local areas,butθiwill vary greatly due to changes in propagation distance over space,resulting in large fluctuations of the received signal power over small distance (i.e.,small-scale fading).Itshould be noted that,ifwe want to make use of the multi-path effect to eliminate the selfinterference signal,the prerequisite is to weak LOS dominated signals(i.e.,passive cancellation method).

To sum up,for the 5 MHz signal,the deep fading region is about 15 cm＜d＜20 cm.The average suppression performance of this region is 15?(?44.12)= 59.12 dB.Therein,our antenna placement accounts for 51.1 dB and deep fading region accounts for 8 dB.Similarly,for the 20 MHz OFDM signal,the deep fading region is about 20 cm＜d＜25 cm.The average suppression performance of this region is 15?(?38.8)=53.8 dB.Therein,our antenna placement accounts for 48 dB and deep fading region accounts for 5.8 dB.The suppression performance of OFDM is worse than the single carrier signal.The reason is that OFDM uses several subcarriers and the bit stream of the information signal is distributed over each subcarrier.The slower symbol rate results in the longer symbol duration time which means it is resistant to the multi-path effect.To the best of our knowledge,itis the firsttime thatpassive cancellation can achieve the value above just using the omni vertical antenna without any auxiliary equipment such as absorber or laptop device[6,11].Amounts of repeated experiment shows thatthe average performance in the deep fading region ofourmethod is betterthan cross-polarization[11,12] as shown in Table 2.

Table 2 Passive suppression measurements

Everett et al.[12]demonstrated that environmental reflections limited the amount of passive self-interference suppression.To measure our method in highly reflective environments,we conduct experiments both in lower reflective environmentand in higher reflective environment. Totally,100 packets were transmitted by the node and the self-interference loss was measured by observing the received signal strength indication(RSSI)from the WARP radio.

As Fig.8 shows,both two kinds of mechanism get worse in highly reflective environment.However,the cross-polarization performance will lose 6.75 dB while our method only lose 3.58 dB on average.The reason is that more reflective environments change the polarization characteristics which decrease polarization isolation ofthe two antennas,thus the suppression performance deteriorate rapidly.As a comparison,the reflective environmenthas a less influence on our method.

Fig.8 Passive cancellation loss comparison

4.1.2 Digitalcancellation

To measure the digitalcancellation performance underdifferent transmit powers,we conductthree experiments using transmit power 5 dBm,10 dBm,15 dBm.The experiment was conducted using 5 MHz QPSK signal.To compare with the method in[2],we select the same data to conductdigitalcancellation one more time.Significant advantages can be seen from Table 3 and Fig.9.It can not only cancel more self-interference power but also has a better performance in the high transmit power.In our opinion,if the transmit power increases,with same passive cancellation,the residualself-interference power will also increase.The proposed digitalcancellation algorithm is easier to compute the temporalstatistics directly ateach time-step to determine the optimalfilter coefficients.

Table 3 Comparison ofdifferent digitalcancellation performances

Fig.9 Digital cancellation versus different transmit power

To show the benefit of our joint mechanism,we select results of the transmit power at 15 dBm as Fig.10 shows. According to our algorithm,the measuredστis around 67.1 ns.Based on those described above,Bc=2.98 MHz and M=3.Fig.10 shows the performance of digital cancellation versus different M andλ.M∈[1,8]and λ∈{0.96,0.98}are selected and taken to the algorithm as the input.From the figure,we can conclude thatour estimation algorithm on M is effective,different M has a significantinfluence on digitalcancellation results butdifferentλhas little influence on it.When M=3,the digital cancellation can achieve the best performance and can cancelabout 26 dB.As Fig.10 shows,M＞3 could notmake performance better.In our opinion,there is no correlation between the front and rear symbols because the random bit sequence is used in our experiment.The correlation between two symbols is only caused by the multipath effect.If the digitalalgorithm imposes correlation beyond the range of multi-path effect,the performance may not get better.As we know,the complexity of the RLS algorithm is O(M2).Our algorithm can use the minimum M to achieve the best performance in the digital domain during each transmission.Therefore,the joint mechanism balances the complexity and performance well.

Fig.10 Digitalcancellation versus different filer length M

To measure the digital cancellation performance,it is also important to focus on how much signal to interference and noise ratio(SINR)willleftafterdigitalcancellation,especially in poor environments.In this paper,SINR is defined as the power of the signalof interestdivided by the sum of the self-interference power and the power of hardware’s noise.The SINR before digital cancellation is defined as SINRB?dcand the SINR after digitalcancellation is defined as SINRA?dc.

As a result,our method is also effective under poor environments as shown in Table 4.Even power of signal of interest is lower than power of self-interference(i.e.,passive and active cancellation methods are not ideal),e.g., SINRB?dc=?4.9 dB,SINRA?dcis still about15.7 dB. However,as SINRB?dcdecreases further,the performance of our method will get worse.This is because of the ADC dynamic range.If the self-interference signal is too high, the quantization error may be introduced when the signal of interest goes through ADC,which may destroy the waveform.Therefore,the digital cancellation mechanism may not only clean the residual self-interference signal, but also cancel some useful signals which will decrease SINRA?dc.

Table 4 Comparison of SINR for digital cancellation method dB

4.1.3 Totalcancellation

To illustrate the benefit of our joint mechanism for total cancellation,we select several previous research results which were based on separate mechanism for comparison. It should be noted thatthe passive cancellation result contains wireless signalattenuation.And the bandwidth ofthe testing signal is 5 MHz.For passive and digital cancellation experiments,if the previous research results are not acquired based on the 5 MHz signal,we willrepeatthe experimentusing the method presented in the priorwork.For analog experiments,we cite the results directly from the previous research because its performance will notsignificant influenced by environments and its high complexity and cost.

Table 5 shows the experimental results.Reference[4] used the antenna separation method and itcan only achieve 29.9 dB.Antenna cancellation was used in[1]as a passive cancellation method.The results shows that it can bring 30 dB extra reduction and the signal will attenuate about 10 dB since the optimal point of the receiving antenna is very close to one of the transmitting antennas.According to[6,13],configuration B in[6]is the best configuration forpassive cancellation.The results shows thatitcan bring 38 dB signal attenuation and the presence of a laptop can make 9 dB impact on the power of self-interference.The directional antennas and cross-polarization and absorber were used in[3,11,12]as a novel passive cancellation method which can achieve 45.5 dB,8 dB,5 dB,respectively.The combination of them can achieve 58.5 dB in our experimentalenvironment.As for the digital method, [1]adopted method in[19]which achieve about10 dB reduction.The rest of the research in Table 5 adopted the frequency-domain channel estimation method to conduct digitalcancellation.This digitalmethod can achieve about 10–20 dB according to our experiments.

Table 5 Totalcancellation performance comparison dB

All the schemes discussed above introduced the analog cancellation method to make sure the totalcancellationenough to bring the self-interference to the noise floor.Reference[1]adopted QHx 220 which can bring 20 dB extra reduction.As for other methods in Table 5,they adopted the same active cancellation method.However,according the conclusion in[13],the presence of the analog cancellation method will make the performance of digitalmethod worse.In fact,the digital cancellation can only achieve about 7 dB after analog cancellation as their papers presented[4,13].As a result,the totalresults with analog cancellation cannotbe acquired by justadding analog cancellation results.In[13],the combination of analog and digital cancellation can achieve about 25 dB reduction.The total cancellation results with analog cancellation can be acquired by adding the passive cancellation performance and 25 dB.

Judging from the overall results,the significant advantage of joint mechanism can be seen from the table and figures.The total cancellation performance can be 85.4 dB which even outperforms other mechanisms with analog cancellation.Meanwhile,it should be noted that if we do not consider the multi-path effect,the result is not better than the method in[3,11,12].

4.1.4 Real-time

There are two main aspects which can affectthe real-time characteristic of a wireless communication system.That is,physical layer performance and media access control (MAC)layerperformance.Forourfull-duplex setup in this paper,only two nodes exchange information from each other.The trigger from the PC controls the nodes transmission/reception,so both of them do notneed to conduct carrier sense and collision avoidance which means there is no extra latency caused by the MAC layer.In the physical layer,we use ttsto denote one time-slot duration. One time-slot is defined the duration between two packets transmission.As Fig.5 shows,the WARPLab mechanism’s workflow and duration of each step are defined as follows:generate IQsamples on a Matlab PC(tDataGen), transfer IQsamples to the boards via ethernet(tEth),transmit/receive samples on board(tDataBoard),received IQ samples transfer back to Matlab PC via ethernet(tEth), process samples on PC(tDataProc).As a result,

The samples read from transmission buffers to DACs and samples written from ADCs to reception buffers are both at 40 MHz.One transmission needs to process about 15 000 samples(payload and control samples)whichThe ethernet and switch we used is 1 Gbps.Thus,tEthis a small and fixed value due to its speed and small length (5 m)and can be negligible.

After 100 experiments,the mean value of tDataGenis 0.015 2 s.Meanwhile,Fig.11 shows the processing time tDataProcfor different filter lengthes on PC.From the result,we can see that even the minimum tDataProc＞tDataGen＞＞tDataBoard.As a result,(6)can be rewritten as t ts≈t Data Gen+t Data Proc.Therein,t Data P roc is obviously a key value for real-time performance of our experiment.The smallerthe tDataProc,the betterthe realtime performance of the system.From Fig.11,tDataProcwill increase quickly(M2)as filter length increases.According to Fig.10 and Fig.11,in certain circumstances, less filter length will bring at least 3 dB loss and more filter length will not bring the significant high cancellation performance but waste longer time.As a result,our jointmechanism which make tDataProcminimum under the premise ofenough cancellation performance.Itshould be noted that the real-time performance bottleneck of our experiment is data generate and process on a Matlab PC. To improve the real-time performance,the useful way is to conductthis step on board which is our future research direction.

Fig.11 Computation time for different filter length

4.2 Comparison with half-duplex

In this part,we are interested in understanding whetherthe proposed self-interference cancellation mechanisms can achieve a full-duplex system thathas a better performance than a half-duplex system.The ergodic rate[13]is the fundamental measure of physical layer capacity in fading channels and is an upperbound on the throughputthatcan be achieved by any MAC protocol.The ergodic rate for transmission to the node i is given bywhere the expected value is computed based on all the packetstransmitted to the node i.And fora full-duplex system,SINR=SINRA?dc.For a two way communication scenario,the sum rate is SR=ER1+ER2.For a fair comparison,the total energy transmitted by a full-duplex node mustbe equivalentto the totalenergy transmitted by a half-duplex node.Thatis

We conductan experimentto evaluate the performance of full-duplex and half-duplex based on 100 packets.The distance between two nodes is 10 m and the channel is a LOS channel.From Fig.12,we can conclude thatthe low power is better for a full-duplex system.This is because the increasing part of the self-interference signal power is higher than the increasing part of the useful signal power atthe receiving node in baseband.According to the experiment results,for transmit power at2 dBm,the gain of the sum rate in our fullsystem is 32%on average.

Fig.12 Full-duplex sum rate compared with half-duplex

Furthermore,we run some simulations based on our experiment results.For LOS channels,our simulations data are calculated based on the data conducted at10 m distance and at 2 dBm transmit power.For the non-LOS channel, our simulations data are calculated based on the data conducted at 3 m distance and at 6 dBm for its deep fading characteristics.

We modelattenuation of the signal of interest and selfinterference using a long-distance path loss model[18]. The pass loss ofthe modelis given by 40+10n lg(D)dB, where D is the distance between communicating nodes. For the LOS channel,n=1.7 and for non-LOS indoor conditions,n=5.

The results of the line-of-channelsimulation in Fig.13 demonstrate thattwo curves intersect at point D≈30 m. Before this point,the sum rate of the full-duplex system is better than the sum rate of the half-duplex system.This range is enough for Wi-Fi links.Result of the non-line-of channel simulation is shown in Fig.14,which indicates thatthe pointof intersection is 6 m.The reason is thatsignalpower is significantly lostover a shortdistance resulting from its severely attenuated property.After 6 m,two curves are almostthe same.

Fig.13 Sum rate comparison between full-duplex and half-duplex versus LOS distance

Fig.14 Sum rate comparison between full-duplex and half-duplex versus non-LOS distance

5.Conclusions

In this paper,we conducta full-duplex system successfully without active cancellation.A new joint passive cancellation method and a new digitalcancellation mechanism are introduced.Itcan estimate the minimum butoptimalfilterlength to conduct digital cancellation which balance performance and computational complexity.Compared with previous methods,the proposed mechanism can achieve 85 dB on average which can bring self-interference to the noise floor.Implementation of a full-duplex prototype is described in ourpaper.Ata 10 m distance with 2–8 dBm transmitpower range,the significantsum rate gain of fullduplex compared with half-duplex can be observed.Meanwhile,itshould be noted that the system presented in this paperis a preliminary study results offull-duplex.In practicalcommunication scenario,itshould consider the channel complexity,mobility,signal bandwidth and the nonreal-time adjustability of antenna position.In the future research,we willconductfurtherstudy and practicaltest.

[1]J.I.Choi,M.Jain,K.Srinivasan,etal.Achieving single channel,fullduplex wireless communication.Proc.of the 6th AnnualInternational Conference on Mobile Computing and Networking,2010:1–12.

[2]M.Jain,J.I.Choi,T.Kim,etal.Practical,real-time,full duplex wireless.Proc.of the 17th Annual International Conference on Mobile Computing and Networking,2011:301–312.

[3]E.Everett,M.Duarte,C.Dick,et al.Empowering full-duplex wireless communication by exploiting directional diversity.Proc.of the 45th Asilomar Conference on Signals,Systems and Computers,2011:2002–2006.

[4]M.Duarte,A.Sabharwal.Full-duplex wireless communications using off-the-shelf radios:feasibility and first results.Proc.of the 44th Asilomar Conference on Signals,Systems and Computers,2010:1558–1562.

[5]M.Duarte,C.Dick,A.Sabharwal.Experiment-driven characterization of full-duplex wireless systems.IEEE Trans.on Wireless Communications,2012,11(12):4296–4307.

[6]A.Sahai,G.Patel,A.Sabharwal.Pushing the limits of full duplex:design and real-time implementation.Houston:Rice University,2011.

[7]M.Duarte,A.Sabharwal,V.Aggarwal,etal.Design and characterization of a full-duplex multiantenna system for wifi networks.IEEE Trans.on Vehicular Technology,2014,63(3): 1160–1177.

[8]D.Bharadia,E.McMilin,S.Katti.Fullduplex radios.Proc.of the ACM Conference on SIGCOMM,2013:375–386.

[9]E.Aryafar,M.A.Khojastepour,K.Sundaresan.Midu:enabling mimo full duplex.Proc.of the 18th Annual International Conference on Mobile Computing and Networking, 2012:257–268.

[10]B.Radunovic,D.Gunawardena,P.Key,et al.Rethinking indoor wireless mesh design:low power,low frequency,fullduplex.Proc.of the 5th IEEE Workshop on Wireless Mesh Networks,2010:1–6.

[11]E.Everett.Full-duplex infrastructure nodes:achieving long range with half-duplex mobiles.Houston:Rice University, 2012.

[12]E.Everett,A.Sahai,A.Sabharwal.Passive self-interference suppression for full-duplex infrastructure nodes.IEEE Trans. on Wireless Communications,2014,13(2):680–694.

[13]M.Duarte.Full-duplex wireless:design,implementation and characterization.Houston:Rice University,2012.

[14]Warplab framework.http://warp.rice.edu/trac/wiki/WARPLab.

[15]C.A.Balanis.Antenna theory:analysis and design.New York:John Wiley&Sons,2005.

[16]G.F.Tu,J.J.Liu,C.Zhang,et al.Studies and advances on jointsource-channel encoding/decoding techniques in flow media communications.Science in China Series F:Information Sciences,2010,53(1):1–17.

[17]J.J.Liu,G.F.Tu,C.Zhang,et al.Joint source and channel decoding for variable length encoded turbo codes.EURASIP Journalon Advances in Signal Processing,2008,(1):7.

[18]T.S.Rappaport.Wireless communications:principles and practice.New Jersey:Prentice Hall,2002.

[19]D.Halperin,T.Anderson,D.Wetherall.Taking the sting out of carrier sense:interference cancellation for wireless LANs. Proc.of the 14th ACM International Conference on Mobile Computing and Networking,2008:339–350.

Biographies

Di Wuwas born in 1989.He received his B.S.degree in information engineering from Beijing Institute of Technology in 2007.He is currently working towards his Ph.D.degree in communication and information systems in University of Chinese Academy of Sciences.His research interests include wireless communication systems,full-duplex and interference cancellation.

E-mail:wudi211@mails.ucas.ac.cn

Can Zhangwas born in 1954.Since July 2003,she has been a professor in the University of Chinese Academy of Sciences.Her research interests include wireless mobile communication,jointsourcechannelcoding/decoding,and information security. E-mail:czhang@ucas.ac.cn

Shaoshuai Gaowas born in 1976.He received his B.S.degree from Tianjin University in 1988,and his Ph.D.degree in signal and information processing from Graduate University of Chinese Academy of Sciences.He did postgraduate research on error resilient video coding in Nanyang Technological University,Singapore from August 2003 to August2004,and he was a researcher in National Institute of Standards and Technology,USA from September 2004 to March 2009.He has been a professor in the University of Chinese Academy of Sciences since May 2009.His main research interests include video processing and communication,wireless sensor networks and physicallayer network coding.

E-mail:ssgao@ucas.ac.cn

Heping Zhaowas born in 1957.He once worked in the State Key Laboratory of Information Security. He is a researcherof China Academy of Space Technology,China Aerospace Science and Technology Corporation.His research interests are secondary planetcommunication and information security.

E-mail:zhpcast@hotmail.com

10.1109/JSEE.2015.00077

Manuscript received July 25,2014.

*Corresponding author.

This work was supported by the National Natural Science Foundation of China(60172045;61032006;61271282).