Practical, Real-time, Full Duplex Wireless - Stanford University

proach cannot provide more than 25 dB of cancellation and cannotbe combined with digital cancellation: it is therefore insufficientfor full duplex.RXControlFeedbackRSSI2.3TXAttenuation& DelayΣFull DuplexMotivated by the above limitations, recent work has proposedantenna placement techniques [4, 5]. The state of the art in full duplex operates on narrowband 5MHz signals with a transmit powerof 0dBm (1mW) [4]. The design achieves this result by augmentingthe digital and analog cancellation schemes described above with anovel form of cancellation called “antenna” cancellation. Figure 2shows a block diagram of this design. The separation betweenthe receive and transmit antennas attenuates the self-interferencesignal [5], but this antenna separation is not enough. The keyidea behind antenna cancellation is to use a second transmit antenna and place it such that the two transmit signals interfere destructively at the receive antenna. It achieves this by having onehalf wavelength distance offset between the two transmit antennas. The resulting destructive interference cancels 20-30dB of selfinterference. Combined with the analog and digital cancellation described above, the design is able to cancel 50-60dB of self-interference,which is sufficient to operate a full duplex 802.15.4 radio.This design, while a breakthrough, has both fundamental andpractical limitations. The first limitation is fundamental and relates to the bandwidth of the transmitted signal. Antenna cancellation ensures that only the signal at the center frequency is perfectly inverted in phase at the receive antenna and thus cancelledfully. However, as a signal moves further away from the centerfrequency, the phase offset of two versions from the transmit antennas shifts away from perfect inversion and so the two do notcancel completely. Correspondingly, antenna cancellation performance degrades as the bandwidth of the signal to cancel increases.In practice, this means antenna cancellation cannot cancel an802.11b signal by more than 30dB: Choi et al. note that it may bejust barely possible to build a full duplex 802.11b node [4]. Widening the signal to 40MHz (e.g., 802.11n) precludes full duplex atWiFi transmit powers.Furthermore, as the cancellation is highly frequency selective,modulation approaches such as OFDM which break a bandwidthinto many smaller parallel channels will perform even more poorly.Due to frequency selectivity, different subcarriers will experiencedrastically different self-interference. Hence cancellation suffersfor wideband OFDM systems that do not adapt on a per subcarrier basis. As WiFi and other wireless technologies move towardsOFDM (802.11n and 802.11g use it) due to its technical advantages, antenna cancellation will encounter serious challenges.Antenna cancellation’s second limitation is requiring three antennas. Full duplex at most doubles throughput. But with threeantennas, a 3x3 MIMO array can theoretically triple throughput.This comparison suggests it may be better to just use MIMO andsacrifice some portion of the available capacity for other useful linklayer primitives such as RTS/CTS. Furthermore, having two transmit antennas creates slight null regions of destructive interferencein the far field.The third limitation relates to the details of Choi et al.’s design.The full duplex radio requires manually tuning the phase and amplitude of the second transmit antenna to maximize cancellation atthe receive antenna. While manual tuning is sufficient for a labproof-of-concept, it is an open question whether it is technicallyfeasible to design a circuit that automatically tunes itself in a dynamic environment.RF ReferenceBalunBalun CancellationBaseband ! RFRF ! BasebandDACADCDigital Interference CancellationFIR filter-DecoderChannelEstimateTX Signal PathDigital InterferenceReferenceEncoderRX Signal PathFigure 3: Block diagram of full duplex system. The ideal cancellation setup uses passive, high precision components for attenuation and delay adjustment.3.DESIGNIn this section we describe the design of a full duplex systemradio that requires only 2 antennas and RF frontends, has no bandwidth constraint, and automatically tunes its self-interference cancellation in response to changing channel conditions.3.1Eliminating the Bandwidth ConstraintThe design is based on a simple observation: any radio thatinverts a signal through adjusting phase will always encounter abandwidth constraint that bounds its maximum cancellation. Tocancel beyond this bound, a radio needs to obtain the perfect inverse of a signal, i.e., a signal which is the perfect negative of thetransmitted signal at all instants. Combining this inverse with thetransmitted signal can in theory completely cancel self-interference.All a radio needs is to invert a signal without adjusting its phase.Luckily, there is a component that does exactly that: a balanced/unbalanced (balun) transformer. Baluns are a common component inRF, audio and video circuits for converting back and forth betweensingle-ended signals – single-wire signals with a common ground– and differential signals – two-wire signals with opposite polarity.For example, converting a single-ended signal on a co-axial cableto a differential signal for transmission on a twisted pair cable (suchas Ethernet), or vice-versa, uses a balun to take the signal as inputand output the signal and its inverse.The key insight of this paper is that one can use a balun in acompletely new way, to obtain the inverse of a self-interferencesignal and use the inverted signal to cancel the interference. Figure 3 shows a 2-antenna full duplex radio design that uses a balun.The transmit antenna transmits the positive signal. To cancel selfinterference, the radio combines the negative signal with its received signal after adjusting the delay and attenuation of the negative signal to match the self-interference.We call this technique balun passive cancellation since we ideally use high precision passive components to realize the variableattenuation and delay in the cancellation path. While balun cancellation can in theory cancel perfectly, there are of course practicallimitations. For example, the transmitted signal on the air experiences attenuation and delay. To obtain perfect cancellation theradio must apply identical attenuation and delay to the inverted sig-

Self-interference path60!240 MHz TX@2.4GHzBalunCancellation (dB)20dBAttenuatorRXVariableVariableAttenuator Delay Line50Balun4030200Cancellation PathFigure 4: Wired setup to measure the cancellation performanceof signal inversion vs phase offset. The phase offset experimentuses an RF splitter instead of a balun to split the signal.Phase Offset20406080Bandwidth in Mhz100120Figure 6: Cancellation performance with increasing signalbandwidth when using the balun method vs using phase offsetcancellation.Received signal (dBm) 60 70Phase Offset 80 90Balun 100 11023002350240024502500Frequency in Mhz25502600Figure 5: Cancellation of the self-interference signal with thebalun vs with phase offset. The received signal is -49dBm without any cancellation. Using a balun gives a flatter cancellationresponse.nal before combining it, which may be hard to achieve in practice.Moreover, the balun may have engineering imperfections, such asleakage or a non-flat frequency response. The rest of this sectionexamines these practical concerns and how to solve them.3.2Balun BenefitsTo understand the practical benefits and limitations of invertinga signal with a balun, compared to inverting it with a phase offset,we conduct a tightly controlled RF experiment. Figure 4 showsthe experimental setup. We program a signal generator to generatea wideband 240MHz chirp with a center frequency of 2.45GHz.This signal goes over two wires. The first wire is an ideal selfinterference path and has a 20dB attenuator representing the antenna separation. The second wire goes through a cancellation path,consisting of a variable attenuator and variable delay element thatcan be controlled to modify the cancellation path signal to matchthe self-interference. The combination of the two signals feeds intoa signal receiver. The variable delay line and attenuator in this experiment are manually tunable passive devices which allow for ahigh degree of precision.We simulate antenna cancellation by making the cancellationpath one half of a wavelength longer than the self-interference path.An RF combiner adds the two signals on the received side to measure the canceled signal. The balun setup, on the other hand, uses abalun to split the transmit signal, and uses wires of the same lengthfor the self-interference and cancellation paths. In both cases, thepassive delay line and attenuator provide fine-grained control tomatch phase and amplitude for the interference and cancellationpaths to maximize cancellation.Figure 5 shows the results. Using a phase-offset signal cancels well over a narrow bandwidth, but is very limited in cancelingwideband signals. Phase offset cancellation can cancel 50dB for a5MHz signal, but only provide 25dB of cancellation for a 100MHzsignal. In comparison, the balun based circuit provides a good degree of cancellation over a much wider bandwidth. For example,balun based cancellation would provide 52dB of cancellation for a5MHz signal and 40dB of cancellation for a 100MHz signal.Balun cancellation is not perfect across the entire band. The keyreason is that the balun circuit is not frequency flat, i.e., differentparts of the band are inverted with different amplitudes. Consequently applying a single attenuation and delay factor to the inverted signal will not cancel the transmitted signal perfectly: thisis a simple instance of real-world engineering tolerances limitingtheory. Based on Figure 5, we can obtain the best possible cancellation with balun and phase-offset cancellation for a given signalbandwidth. Figure 6 shows the best cancellation achieved usingeach method for signals of varying bandwidths.3.3Auto-tuningThe results in Figure 6 show that, if the phase and amplitude ofthe inverted signal are set correctly, balun cancellation can haveimpressive results across a wide bandwidth. This raises a simple follow-on question. Is it possible to automatically adjust thephase and amplitude, thereby self-tuning cancellation in responseto channel changes? In this section we describe an algorithm thatcan accurately and quickly self-tune a cancellation circuit.The basic approach is to estimate the attenuation and delay of theself-interference signal and match the inverse signal appropriately.Ideally, the auto-tuning algorithm would adjust the attenuation anddelay such that the residual energy after balun cancellation wouldbe minimized (assuming no other signal is being received on theRX antenna). Let g and τ be the variable attenuation and delayfactors respectively, and s(t) be the signal received at the input ofthe programmable delay and attenuation circuit. The delay overthe air (wireless channel) relative to the programmable delay is τa .The attenuation over the wireless channel is ga . The energy of theresidual signal after balun cancellation is:ZE (ga s(t τa ) gs(t τ ))2 dt(1)Towhere To is the baseband symbol duration. The goal of the algorithm is to adjust the parameters g and τ to minimize the energy ofthe residual signal.Our insight is that the residual energy function in Eq. 1 has apseudo-convex relationship with g and τ for WiFi style OFDM signals. We omit the mathematics for brevity but note that we can exploit this structure to design a simple gradient descent algorithm toconverge to the optimal delay and attenuation.3.3.1Practical Algorithm with QHx220While a gradient descent algorithm would work well, fine-grained

RXTXAir SignalQhx220RSSIOutput gaingigaingqFixedDelay(τ) Wire SignalAlgorithmBalunSignal toTransmitControl gi, gqRF Interference CancellationCircuitFigure 7: Block diagram of full duplex system with Balun Active Cancellation. The RSSI values represent the energy remaining after cancellation. Our algorithm adapts attenuationparameters (gi and gq ) to minimize this energy.programmable analog attenuation and delay lines are unfortunatelynot typical commodity components and so are expensive. Hencein our current implementation we have to make do with a component that provides an approximation, the QHx220 noise cancellation chip [11], which prior full duplex designs have used [4, 12].Since the QHx220 is an active component, we call this cancellationscheme Balun Active Cancellation.Figure 7 shows the block diagram of Balun Active Cancellationwith the auto-tuning circuit. The RSSI value provides the residualsignal energy after balun cancellation has subtracted self interference from the received signal. As we can see, the QHx220 doesnot actually provide a variable delay. Instead, it takes the inputsignal and separates it into an in-phase and quadrature component.The quadrature component has a fixed delay (τ ) with respect to thein-phase component. It emulates a variable delay by controllingthe attenuation of the in-phase and quadrature signals (gi and gq ),adding them to create the output.For a single frequency, this approach can correctly emulate anyphase. However, for signals with a bandwidth, the fixed delay τonly matches one frequency. This fixed delay suffers from a similarbandwidth constraint as antenna cancellation. Prior work also discusses the limitation of the cancellation model used by QHx220 [9].The goal of the auto-tuning algorithm in this case would be tofind the attenuation factors on both lines such that the QHx220 chipoutput is the best approximation of the self interference we needto cancel from the received signal. Fortunately, we can show thateven with this approximate version, we still retain a psuedo-convexstructure. To see this empirically, we conduct an experiment wherethe TX antenna transmits a 10MHz OFDM signal, and we vary thetwo attenuations in QHx220. We plot the RSSI output in Figure 8,where a deep null exists at the optimal point. Hence we can usethe same gradient descent algorithm for tuning the two attenuationfactors in QHx220, gi and gq . The algorithm works in steps, andat each step it computes the slope of the residual RSSI curve bychanging gi and gq by a fixed step size. If the new residual RSSI islower than before, then it moves to the new settings for the attenuation factors, and repeats the process. If at any point it finds thatthe residual RSSI increases, it knows that it is close to the optimalpoint. It then reverses direction, reduces the step size and attemptsto converge to the optimal point. The algorithm also checks forfalse positives, caused due to noisy minimas. The algorithm is fast;it typically converges to the minimum in 8 15 iterations, depending on the choice of the starting point. An iteration requires 5 measurements, each involving a phase/attenuation adjustment followedby RSSI sampling.Figure 8: RSSI of the residual signal after balun cancellationas we vary gi and gq in the QHx220. Note the deep null at theoptimal point.Caveats: The commodity parts of our prototype introduce a fewlimitations. The QHx220 has hardware constraints, such that itcauses non-linear distortion, especially for input powers beyond 40dBm. Hence, cancellation will not be perfect for typical wirelessinput powers (0-30dBm). Non-linear distortion also impacts digitalcancellation, as we see in Section 5.1.Our current implementation uses the QHx220 despite its imperfections because it is inexpensive and easily available. However,we believe that it is feasible to build a full duplex radio using anelectronically tunable delay and attenuation chipset, since they arecommercially available but just not widely and inexpensively. Furthermore, including them as small parts of a full duplex radio hardware design would not be particularly complex or expensive.3.4Digital CancellationAs Figure 5 shows, a industry-grade balun cancellation circuitcan cancel up to 45dB of a 40MHz signal. But this cancellationonly handles the dominant self-interference component betweenthe receive and transmit antennas. A node’s self-interference mayhave other multipath components, which, although much weakerthan the dominant one, are strong enough to interfere with reception. Furthermore, the balun circuit may distort the cancellationsignal slightly, such that it introduces some interference leakage.The full duplex radio design uses digital cancellation (DC) tocancel any residual interference that persists after balun cancellation. Implementing DC for a full duplex radio, however, is morechallenging than other uses of digital cancellation, such as successive interference cancellation (SIC) [8] and ZigZag decoding [7].Unlike SIC or ZigZag, which use DC to recover packets whichwould have otherwise been lost, a full duplex radio uses DC to prevent the loss of packets which a half duplex radio could receive.While an SIC implementation that recovers 80% of otherwise lostpackets is a tremendous success, a full duplex radio that drops 20%of packets is barely usable.To the best of our knowledge, our DC system has three novelachievements compared to existing software radio implementationsin the literature. First, it is the first real-time cancellation implementation that runs in hardware: this is necessary for the MACexperiments in Section 5.2. Second, it is the first cancellation implementation that can operate on 10MHz signals. Finally, it is thefirst digital cancellation that operates on OFDM signals. For thesereasons, we give a very detailed description of its design and algorithms.Digital cancellation has two components: estimating the selfinterference channel; and using the channel estimate on the knowntransmit signal to generate digital samples to subtract from the received signal.

Channel Estimation: To estimate the channel, the radio uses knowntraining symbols at the start of a transmitted OFDM packet. It models the combination of the wireless channel and cancellation circuitry effects together as a single self-interference channel, estimating its response. The estimation uses the least square algorithm [15]due to its low complexity. Since the training

2. RADIO DESIGN AND FULL DUPLEX This section provides background on the basics of radio design as well as full duplex. It explains the basic challenges in building a full duplex radio, the existing techniques for doing so, and the limitations of those techniques. 2.1 Radio Design Figure 1 shows the basic design of a modern radio receiver. We