Deep Neural Network a Posteriori Probability Detector for Two-dimensional Magnetic Recording

Primary author: Jinlu Shen
Faculty sponsor: Benjamin Belzer, Krishnamoorthy Sivakumar

Primary college/unit: Voiland College of Engineering and Architecture
Campus: Pullman

Abstract:

The magnetic recording channel in hard disk drives is a binary inter-symbol interference (ISI) channel that typically adopts a linear minimum mean square error (MMSE) equalizer with partial response (PR) signaling followed by a trellis-based detector such as Bahl-Cocke-Jelinek-Raviv (BCJR) or Viterbi. In two-dimensional magnetic recording (TDMR), an array of heads read data from multiple adjacent tracks in order to equalize inter-track interference (ITI), which is severe in high density recording. The multi-track effects combined with pattern-dependent noise inherent to HDD recording channels lead to a trellis state explosion problem, when an auto-regressive model is used for pattern dependent noise prediction (PDNP). The detector complexity grows exponentially with ISI channel length I and noise predictor order L, and becomes impractical for more than two tracks.
As a solution, we propose a novel deep neural network (DNN). The DNN detector replaces the typical Viterbi-PDNP or BCJR-PDNP, directly outputs log likelihood ratios of the coded bits and iteratively exchanges them with a channel decoder to minimize decoded BER. Three DNN architectures are investigated – fully connected DNN, convolutional neural networks (CNN), and long short-term memory (LSTM). The DNN’s complexity is limited by employing MMSE equalizer pre-processing. The best performing DNN architecture, CNN, is selected for iterative decoding with a channel decoder. Simulation results on a realistic media model shows as much as 30.47% detector BER reduction, and as much as 21.72% areal density gain compared to a conventional system.