Methods for Non-linear Distortion Immune End-to-End Learning with Autoencoder-OFDM


Reference #: 01441

The University of South Carolina is offering licensing opportunities for Methods for Non-linear Distortion Immune End-to-End Learning with Autoencoder-OFDM1


AI-based communication systems utilize machine learning modules (e.g., deep learning) to replace the functionality of the highly engineered blocks (e.g., coding, modulation, waveform, etc) in the physical layer of communication systems. However, the machine learning methods in the field of computer vision may not be directly applied to the communication systems as a communication system may encounter different challenges. One of the challenges related to communication systems is the high instantaneous power fluctuations. Although there are some methods that limit the instantaneous peak power in the literature, these methods may require further training to overcome the non-linearities. In this invention, we solve the instantaneous peak problem of an AI-based communication system by introducing a new layer without any training, i.e., Golay layer.

Invention Description:

This technology provides a new layer (Golay layer) tailored for AI-based communication systems to limit the instantaneous peak power for the signals. The disclosed methods can also be applied to communication devices that operate under power-limited link budgets while autonomously decreasing the error rate for auto-encoder orthogonal frequency division multiplexing. The invention may be part of a wireless standard (e.g., 5G and beyond, IEEE 802.11 WLAN) that allows AI-based communication systems. The disclosed method may also decrease the training complexity. The embodiment relies on the manipulation of the complementary sequences through neural networks. We disclose how to stabilize the mean power and peak power by using the algebraic representation of the complementary sequences. The introduced layer may be used with several other basic layers, such as the clipping layer and polar-to-cartesian layer as the Golay layer operates in the polar coordinate. By also changing the parameter of the Golay layer, it can also allow constant-amplitude sequence in the frequency domain.

Potential Applications:

Traditional end-to-end learning (e.g., auto-encoder orthogonal frequency division multiplexing (AE-OFDM)) methods do not provide a permanent solution for PAPR without a rigorous training/optimization procedure, which may potentially increase the training complexity in practice. Solutions that control instantaneous power fluctuations are needed for artificial intelligence (AI) based transmitter and receivers to decrease the training complexity. This invention addresses this problem.

Advantages and Benefits:

The disclosed methods can also be applied to communication devices that operate under power-limited link budgets while autonomously optimizing its error rate performance for auto-encoder OFDM. The disclosed method may decrease the training duration as the transmitter and receiver do not deal with the PAPR problem with the Golay layer. The Golay layer itself ensures the low PAPR.

Patent Information:
For Information, Contact:
Technology Commercialization
University of South Carolina
David Matolak
Alphan Sahin
Complementary sequences
machine learning
© 2024. All Rights Reserved. Powered by Inteum