From the previous section it is clear that multiple stages of cancellation are needed. They can be categorized into digital self-interference cancellation (SIC) and analog SIC. The latter is split into cancellation techniques at RF and at baseband (BB), as shown in Figure4.4. RF SIC is the most crucial for the LNA, as the cancellation ensures that the LNA does not saturate. The combination of RF and BB self-interference cancellation ensure enough dynamic range on the ADC. Finally, digital SIC cleans up any remaining SI. The different steps are visualized in Figure4.2. Before analog cancellation (Figure4.2c), only the SI is visible, after analog cancellation (Figure4.2d), the useful signal becomes visible, although heavily distorted. Finally, after digital cancellation (Figure
4.2e) the useful signal is clearly visible.
4.2.1
Analog self-interference cancellation
Several research groups have proven the feasibility of analog SIC. Let us first go over the techniques at RF. These techniques typically focus on reducing the
main component of the SI signal (cfr. Figure4.3). If more than one antenna is used, this can be achieved by separating the antennas in such a way that the two transmit antennas create destructive interference at the receiver antenna [35]. If precisely placed, this allows to cancel the SI up to 30 dB. However, this only works for a specific frequency and antennas have to be moved to use another frequency. Another technique which uses antennas, is to use dual polarizing antennas [36]. One polarization is used for the transmitter and the other is used for the receiver. The isolation between the two polarizations can be up to 40 dB [36]. The main drawback is that if two devices communicate with each other, their polarization need to be aligned for best SNR.
Before going to techniques using custom hardware, there is one technique which uses existing hardware in a wireless transceiver to cancel the SI. By using an extra transmitter chain it is possible to inject an inverse signal of the SI which destructively interferes with the leaked self-transmitted signal [37]. The idea is to pre-compensate for the distortions in the first transmit chain. Of course the second transmit chain adds distortions of its own which should also be calibrated for. In [37] it is shown that such a technique can achieve a cancellation of up to 23 dB. However, adding a second transmitter chain requires more space on the chip as well as more power. Ultimately the benefits would probably be lower than using the two transmitter chains in a multiple input multiple output (MIMO) configuration.
The techniques that follow in the rest of this section all use custom hardware designed for IBFD operation. The first one uses an electrical balance duplexer (EBD) [38], of which the concept is shown in Figure4.5. The EBD creates two signal paths for the signal coming from the power amplifier (PA), one part goes to the antenna and is reflected back towards the LNA, another part goes to the balance network and also gets reflected towards to LNA. The reflection from the antenna is due to mismatch of the antenna and causes the main linear components of the SI signal. To achieve cancellation, the balance network, which consists of tunable capacitor banks, should be set to create an inverse copy of the reflected signal from the antenna port. These two signals will add up before the LNA and will destructively interfere. If correctly tuned, such architectures can achieve up to 70 dB of cancellation over a 5 MHz bandwidth and up to 50 dB over a 20 MHz bandwidth. On top of this, in [38] it was shown that these devices can be made very linear, lowering the non-linear and noise components and therefore relaxing the subsequent cancellation steps. However, tuning of the balance network is key, any changes in the antenna reflection should be matched by changes in the balance network. The speed and accuracy at which this can be achieved determines the performance in practical situations. Moreover, due to the signal split at the PA side, there is a 3 dB insertion loss inherent to this architecture, which should be compensated.
STATE-OF-THE-ART SI CANCELLATION TECHNIQUES 49
Figure 4.5: The electrical balance duplexer creates two signal paths. At the receiver end, these two signals destructively interfere with each other the cancel to self-interference.
As the linear components of the SI signal are a linear combination of delayed versions of the transmitted signal, it is possible to create a similar path for the cancellation signal. In [11], the signal from the transmit chain is tapped and passed through a network of parallel fixed lines of varying delay and tunable attenuators. These copies are then added up again and subtracted from the received signal. The cancellation is limited by the range of the varying delay lines. Moreover, adding more lines is typically limited by space or power constraints. Still, this technique allows to cancel the SI signal up to 45 dB with careful tuning [11]. The main downside is that this techniques require space on chip for the delay lines, which can be costly.
Finally, at baseband, a vector modulator is a component which can mix a certain signal down to baseband and at the same time apply a variable phase shift and attenuation [39]. The input of the vector modulator is taken at RF while the output is added with the signal coming out of the mixer in the receiver chain. The phase shift and attenuation is adjusted in such a way that both signals destructively interfere with each other. In [39] it is reported that up to 27 dB of cancellation can be achieved. Another vector modulator design taps the signal at the transmitter baseband and then mixes the signal to add a variable delay and attenuation [40]. Their design is able to cancel up to 42 dB, although with very bad linearity specifications. These techniques can be added to all the previous techniques to cancel the SI before the ADC.
4.2.2
Digital self-interference cancellation
In the digital domain, there is more flexibility to cancel the SI signal and clean up any remaining linear and non-linear components. To cancel the linear components, the remaining self-interference channel (h11 in Figure4.1)
should be estimated using for example a least squares estimator. The received self-interference can be written in the frequency domain as
Y = H11X1+ N. (4.2) The channel H11 can therefore be estimated using
ˆ H11 = Y
X1, (4.3) as Y are just the received samples and X1 are the known transmitted samples.
This estimation can be done on a per-packet basis using the preamble which is available in most wireless standards. Moreover, most of the hardware necessary for these operations is already available in a typical receiver as it is required to compensate for any channel variations.
Unfortunately, the rather simple channel estimation from Equation (4.3) only takes the linear components into account. To cancel the non-linear components, more computations are needed. In [11] a Taylor expansion is used to approximate the non-linear channel. Therefore, the received digital samples can be written as,
y(n)=X
m
x(n)(|x(n)|)m−1∗ h
m(n), (4.4)
where hm(n) are the terms that need to be estimated. Empirically, in [11] it
was found that most higher order (m) terms are zero. This comes from the fact that higher order terms are created by mixing lower order terms together, and therefore the power in the higher order terms reduce. Therefore, in a typical indoor environment only 224 hm(n) values need to be estimated, which
is something that can again be done on the preamble.
After estimating both the linear and non-linear components, the resulting channel is applied on X1 and subtracted from the received samples, giving
Yd = (H11− ˆH11)X1+ N. (4.5)
CURRENT USE-CASES FOR IN-BAND FULL DUPLEX 51
Figure 4.6: The packet transmissions using in-band full duplex relaying show the overlapping transmissions. (Source: [41])