Recurrent neural network equalization for partial response shaping of magneto-optical readback signals

Abstract

In this paper, use of recurrent neural network equalizer (RNNE) in place of linear equalizer (LE) to combat both linear and nonlinear distortions corrupting the Magneto-optical (MO) readback signal is discussed. It is shown that RNNE can outperform LE without introducing significant complexity. RNNE is used to equalize the MO recording readback signal corrupted by transition jitter, intersymbol interference (ISI) and additive white Gaussian Noise (AWGN) at a density of 50 kbpi. The MO signal is equalized to a, partial response (PR) (1 + D) using either the RNNE or the LE and the equalizer's mean-squared-error (MSE) performance is compared. Then, the equalized signal is passed through a detector and it is shown that a signal equalized to a PR (1 + D) shape can be detected using either a bit-by-bit type of detector (BD) or a sequence detector implemented via Viterbi Algorithm (VA). The bir-error-rate (BER) performance of ED is compared to that of the Viterbi detector and it is shown that PR equalization of MO readback signals using RNNE improves MSE performance over linear equalizer, allowing use of ED rather than LE+Viterbi Algorithm with comparable BERs.