Block-Structured Deep Learning-Based OFDM Channel Equalization

This letter considers equalization of an optical signal under the Orthogonal Frequency Division Multiplexing modulation scheme. The equalizer mitigates nonlinear effects caused by a power amplifier (PA) in the transmitter and a transimpedance amplifier (TIA) in the receiver. We compare the Convolutional Neural Network (CNN) and the Wiener-Hammerstein (WH) linearization models using two PA and three TIA devices. We are first to demonstrate a great boost in CNN effectiveness if sequentially combined with a linear equalization. The optimal sequence of linear and nonlinear blocks depends on the device profiles, and is found to be the same between CNN and WH. We develop a new block-structured CNN-based solution that utilizes the optimal sequence and brings up to 2.88 dB Q-factor gain over the traditional WH approach.

Automated summary

This letter discusses the equalization of an optical signal under the Orthogonal Frequency Division Multiplexing modulation scheme, addressing the nonlinear effects caused by a power amplifier (PA) and a transimpedance amplifier (TIA). Comparisons are made between the Convolutional Neural Network (CNN) and the Wiener-Hammerstein (WH) linearization models, using multiple PA and TIA devices. The study finds that combining CNN with linear equalization leads to a significant improvement in effectiveness. Furthermore, a new block-structured CNN-based solution is developed, incorporating the optimal sequence and providing up to a 2.88 dB Q-factor gain over the traditional WH approach.