A Study of Interference Suppression Using Deep Learning Methods

Abstract

This thesis investigates a Deep Learning model for interference suppression in wireless communications. By exploiting the structure of Convolutional Neural Network-based autoencoders, we develop an approach for interference suppression with no prior knowledge on characteristics or the exact location of interference. Traditional interference suppression techniques are heavily reliant on specific domain knowledge, thus their applicability in dynamic wireless environments is limited. This thesis proposes a CNN-AE (Convolutional Neural Network - Autoencoder) model that consists of an encoder, which captures the latent space representation from the input data, and a decoder that reconstructs the desired signal to suppress interference effects. We investigate the performance of a QPSK-based wireless communication system with three explicit interference scenarios, namely, %in-band tone, out-of-band tone, single frequency tone interference with two cases of in-bandwidth and out-of-bandwidth, and wideband interference from a dataset that captured over the air communication signals. A study is performed for different SNR values along with the SINR values to observe the effectiveness of the approach at different levels. The results of our approach are quantified using popular metrics such as bit error rate (BER), error vector magnitude (EVM), and Signal to Noise-Interference Ratio (SINR). The proposed model outperforms the baselines with classical techniques such as matched filtering and least squares adaptive filtering consistently over these several metrics. The thesis also investigates the latent space behavior of the autoencoder; which is used to provide an interpretation of how the network classifies between the desired signal and interference. We use this contextual information to pursue future directions in interference suppression performance by exploiting cyclostationarity properties of our desired signal to our advantage. One of the important contributions in this work involves carrying out thorough analysis with respect to the generalization capability of CNN-AE for different types of interference and signal conditions. The results presented herein illustrate the potential of a deep learning-based approach in enabling more robust and adaptive wireless communication systems that would be capable of autonomously managing complex interference scenarios without human intervention.