Architecture Optimization of a CNN Media Noise Estimator for TDMR

We investigate several architectures for a convolutional neural network (CNN) media noise estimator (MNE) that sends media noise estimates to a three-track Bahl–Cocke–Jelenik–Raviv (BCJR) 2-D magnetic recording (TDMR) detector. While novel CNN architectures have previously been proposed for image classification problems, here the CNN estimates media noise based on inputs from the BCJR detector and from a linear partial response equalizer. The CNN architecture used in our previous work on CNN-based TDMR detection is used as a comparison baseline. Experiments show that an optimized CNN MNE with three residual path connections achieves up to a 3.2% reduction in mean squared estimation error, a 3.7% detector bit error rate (BER) improvement, and a 71.9% reduction in computational complexity compared to the baseline.