Blind Channel Estimation and Deep-Learning-Based Self-Interference Mitigation in Full-Duplex Relay Links

We analyze performance of two-hop full-duplex relay system in the presence of residual self-interference (SI) and frequency-selective fading. In particular, feasibility of the second-order statistics-based blind channel estimation in the context of two-hop full-duplex relay systems is investigated. As part of this process, we evaluate blind estimation of the source-to-relay (SR), relay-to-destination (RD), and source-to-destination (SD) channels. To that end, the performance of blind and traditional pilot-based (training sequence based) channel estimation approaches are compared. This is accomplished by deriving the Cramer-Rao Lower Bound (CRLB) expressions for both blind and pilot-based schemes, and comparing them to each other and to the mean-squared error (MSE) values measured via simulation. Post-equalization SINR expressions are also derived for both blind and pilot-based methods. Furthermore, a modified post-equalization SINR expression, where the channel estimation error is replaced by the inverse of the Fisher information matrix (FIM) is proposed, providing an upper bound for the post-equalization SINR. These analytically-predicted SINR values for the blind approach are compared to the SINR measured via link simulation. The performance of the two channel estimation methods is analyzed by comparing the CRLB, post-equalization SINR, and BER performance for two significantly different transmission packet sizes. All of the above metrics indicate that blind estimation provides clear performance advantages relative to the pilot-based counterpart. Additionally, the blind approach eliminates the overhead associated with the pilot-based method, where a portion of the system resources is allocated to the pilot sequence. To quantify this, the spectral efficiency of the FD relay system employing blind and pilot-based channel estimation methods are compared, indicating that at high SNR, the blind approach provides around 2-bps/Hz spectral efficiency gain relative to a typical pilot-based method. The computational complexity of the two channel estimation techniques are evaluated and compared. The blind approach has a clear computational advantage for larger packet sizes. Additionally, we analyze the performance of a full-duplex relay system in the presence of purely linear residual self-interference and frequency-selective fading. In particular, the residual SI channel estimation performance is evaluated both analytically and via simulation. The BER performance at the destination is also characterized via simulation. Two schemes are considered for the equalization of the source-to-relay (SR) and relay-to-destination (RD) channels: end-point equalization and distributed equalization. In the first approach, channel equalization is performed only at the destination and is similar to an amplify-and-forward model. In the second approach, equalization of the SR channel is performed at the relay, while the RD channel equalization is performed at the destination. We show that at the cost of a modest complexity increase at the relay, the distributed scheme provides more robust performance than its end-point counterpart. We also propose and evaluate a new method for digital mitigation of the non-linear self-interference in the context of a full-duplex relay link. The proposed method utilizes Deep Neural Networks (DNN) for reconstruction of the non-linear transmitted signal. Most known DNN-based SI mitigation techniques require a large number of on-line training samples taking the resources away from actual data communication. Our approach allows complete elimination of on-line training by decoupling estimation of the linear SI propagation channel from reconstruction of the non-linear transmitted SI signal. We utilize link-level simulation to demonstrate the SI cancellation performance of the proposed algorithm comparing it to the existing state of the art DNN-based solutions. The simulation results also confirm that robust BER performance can be achieved based on the proposed SI mitigation technique for SIR levels as low as -20dB. Generalization error in the context of a feed-forward DNN is attributed to an insufficient coverage by the training set, resulting in a discrepancy between the statistics of the training set and those of the run-time data. We derive an upper bound for the generalization error as a function of the training set size and compare the derived bound to measured generalization error values. Additionally, an analytical approximation expression for the probability of the generalization error is proposed and evaluated against the simulation results. This comparison indicates that the proposed expression provides a sufficiently close approximation of the measured probability of generalization error and, hence, can be used as a guidance in selecting the appropriate number of training samples. This provides an alternative to the trial and error approach in selecting the size of the training set.