Design of Deep Learning-Based 1-bit Transceiver With Oversampling and Faster-Than-Nyquist (FTN) Signaling

The adoption of one-bit analog-to-digital converters (ADCs) in wireless communication has become increasingly popular, owing to their potential to reduce power consumption. However, there is a decreased spectral efficiency (SE) associated with their use, as they are unable to take advantage of high-modulation-order signaling. Faster-than-Nyquist (FTN) signaling along with oversampling are promising means of improving SE at the expense of additional inter-symbol interference (ISI). This study establishes a common platform for design and evaluation by providing a general system model and problem formulation, as well as proposing an alternative solution framework for the one-bit quantization system. By systematically evaluating optimizable factors that impact performance, such as signaling rate, transmit sequences, and error correction blocks, this study proposes a deep learning (DL)-based architecture, specifically channel autoencoders, for end-to-end communication over a one-bit quantization channel. Various transceiver designs are proposed for performance comparison as well as to enable one-bit quantized communication at previously unattainable information rates through conventional means. Numerical results showcase the superiority of jointly optimizing all blocks in an additive white Gaussian noise (AWGN) channel. The DL-based scheme operationalizes Bit Error Rate (BER) performance at information rates scaling up to 80% of Shamai’s limit. Beyond AWGN, the autoencoder-based transceiver design extends to make one-bit quantization viable over a challenging Rayleigh multipath fading channel. A detailed analysis compares a pilot-based scheme for one-bit quantization with a pilotless option, thereby revealing their SE and BER relationship. The pilot-based quantized scheme outperforms conventional fading channel transmission without quantization by up to 10 dB.