AI-based Equalization of Faster-than-Nyquist Multicarrier Signaling

This thesis investigates advanced signaling and receiver design for high-efficiency non-orthogonal communication systems, focusing on Multicarrier Faster-than-Nyquist (FTN) signaling. To overcome the severe Inter-Symbol Interference (ISI) and Inter-Carrier Interference (ICI) induced by spectrally aggressive time-frequency compression, a deep learning-based architecture utilizing a specialized 1D/2D Residual Network (ResNet) is proposed. The framework incorporates a Virtual Padding (VP) strategy to dynamically handle channel memory boundaries under realistic operational constraints. The proposed deep neural network explicitly learns the underlying interference geometries offline, removing the need for real-time algebraic injection of the exact channel Gram matrix required by classical benchmarks like the iterative ID+FSD receiver. Numerical results demonstrate that the ResNet architecture achieves near-optimal pragmatic capacity limits, particularly in multi-dimensional 2D packing scenarios where the network effectively converts dense bidirectional cross-correlations into useful contextual information.