Automatic adaptive wireless demodulator using incremental learning in real time

Abstract

In wireless communication systems, a received signal is corrupted by various means, such as noise, multi-path fading, and defects in hardware. To properly demodulate the signal and recover information, complex systems are used. This typically consists of a series of filtering, corrections, timing recovery, and finally demodulation. Furthermore, the approaches for each stage are application specific. Deep learning (DL) can be applied to create an automatic demodulator, independent of modulation type, with no preprocessing, replacing the complex traditional system. However, these systems can only handle scenarios that are incorporated at the initial training stage. If new modulation types are encountered, the system must be re-trained to adapt. Traditional DL systems require the entire original dataset to retain old information, which increases storage requirements and training time. To increase adaptability, we incorporate incremental learning (IL) into a DL demodulator. Incremental learning attempts to overcome these issues, allowing a system to train on only new information. We apply IL to learn to demodulate new modulation types, not initially introduced to this system. We demonstrate this system in the field through the use of software defined radio. The system is subjected to unknown modulation types, and shown to adapt in real-time and over-the-air in an unsupervised environment.