As wireless communications become increasingly synonymous with everyday life, the demand for higher data rates, reliability, and efficiency continues to grow. This is further accelerated by the rapid rise in the Internet of Things (IoT) and industrial automation. However, traditional algorithm-based signal processing is limited by algorithmic complexity and its limited ability to adapt to and cope with increasingly adverse and congested channel conditions, thereby reducing the effectiveness of traditional digital signal processing techniques in real-world environments. To address these challenges, approaches using Deep Learning (DL) have rapidly gained attention as a promising alternative to traditional DSP techniques. DL techniques excel in adaptability and have demonstrated on-par or even superior performance compared to traditional approaches for various RF environments, particularly in challenging conditions such as low-SNR, high-mobility, and non-ideal channel scenarios. In this survey, we examine the various stages that comprise popular wireless transmission techniques, specifically Orthogonal Frequency Division Multiplexing (OFDM), which underpins numerous technologies, including Wi-Fi, 4G LTE, 5G, and DVB. We review recent research activities to implement the various stages of the OFDM receiver chain using DL methods, including synchronization, Cyclic Prefix (CP) removal, Fast Fourier Transform (FFT), channel estimation and equalization, demodulation, and decoding. We also review approaches that take a holistic view, aiming to use a unified DL approach across the entire signal processing chain. For each stage, we review existing Deep Learning-based methods and provide insights into how they aim to meet or exceed the performance of traditional approaches. This survey provides a comprehensive overview of the current development of deep learning-based OFDM systems, highlighting the potential benefits and challenges that remain in fully replacing conventional signal processing methods with modern deep learning approaches.
Bray et al. (Fri,) studied this question.