Serial and parallel convolutional neural network schemes for NFDM signals.

Area of Science:

• Optical Communications
• Machine Learning
• Signal Processing

Background:

• Nonlinear frequency division multiplexing (NFDM) is crucial for high-capacity optical networks.
• Directly decoding NFDM signals using machine learning remains a challenge.
• Hardware implementation considerations are vital for practical applications.

Purpose of the Study:

• To propose and analyze two conceptual convolutional neural network (CNN) schemes for direct NFDM signal decoding.
• To evaluate the feasibility of CNNs for practical NFDM-based fiber optic communication.
• To consider hardware implementation aspects for both serial and parallel network designs.

Main Methods:

• Development and analysis of two CNN architectures: a serial network for smaller applications and a parallel network for high-speed data centers.
• Training both network schemes using simulated NFDM data.
• Evaluating network performance based on accuracy and memory footprint.

Main Results:

• The serial CNN scheme requires only 0.5 MB of memory, suitable for small-scale applications.
• The parallel CNN scheme utilizes 128 MB of memory, enabling high-speed parallel computing.
• Both proposed CNN schemes achieved decoding accuracy exceeding 99.9%.

Conclusions:

• CNNs demonstrate significant potential for practical and accurate decoding of NFDM signals in optical communication systems.
• The developed serial and parallel CNN architectures offer versatile solutions for different application requirements, from small users to data centers.
• The high accuracy achieved validates the effectiveness of CNN-based approaches for advanced optical signal processing.