Semi-Supervised Feature-Crosses Neural Network Equalizer in Fiber Optics

In this paper, a semi-supervised feature-crosses nonlinear equalization(FC-NLE) scheme with adaptive threshold regulation and consistency regularization is proposed to compensate for the linear and nonlinear impairment in the high-speed optical communication systems. The feasibility of the proposed method is verified in a 120 Gb/s quadratic pulse amplitude modulation (PAM4) transmission based on intensity modulation and direct detection (IM/DD). After a 5 km transmission, the semi-supervised FC-NLE scheme enables the bit error rate(BER) to reach the hard-decision forward-error-correction(HD-FEC) threshold of 7% when the received optical power is more than 4.7 dBm with only 4% proportion of labeled data required, which has more than 1.3 dB and 1dB better than that of FC-NLE and the traditional transductive learning semi-supervised neural network scheme.