Deep Learning-based Turbo-detection and Equalization for Two- and Three-dimensional Magnetic Recording

Amirhossein Sayyafan

doi:10.7273/000005107

This dissertation considers various machine-learning based signal processing architectures for equalization and detection of two- and three-dimensional magnetic recording signals for hard disk drives. The first part of this dissertation (Chapter 2) investigates a combined Bahl-Cocke-Jelinek-Raviv (BCJR) and deep neural network (DNN) turbo-detection architecture for one-dimensional (1D) hard disk drive (HDD) magnetic recording. Simulated HDD readings based on a grain flipping probabilistic (GFP) model are input to a linear filter equalizer with a 1D partial response (PR) target. The equalizer output is provided to the BCJR detector in order to minimize the intersymbol interference (ISI) due to the PR mask. The BCJR detector's log-likelihood-ratio (LLR) outputs (along with the linear equalizer outputs) are then input to the DNN estimator, which estimates the signal dependent media noise. The media noise estimate is then fed back to the BCJR detector in an iterative manner. Several DNN media noise estimation architectures based on fully connected (FC) and convolutional neural networks (CNNs) are investigated. For GFP data at 48 nm track pitch and 11 nm bit length the CNN-based BCJR-DNN turbo detector multiplies the detector bit error rate (BER) by 0.334× compared to a BCJR detector that incorporates 1D pattern-dependent noise prediction (PDNP). Chapter 2 also considers a concatenated BCJR detector, LDPC decoder and DNN architecture for a turbo-detection system with 1D and 2D magnetic recording (1DMR and TDMR). The input readings first are fed to a PR equalizer. Two types of equalizer are investigated: a linear filter equalizer with a 1D/2D partial PR target and a CNN PR equalizer. The equalized outputs are passed to the BCJR to generate the LLR outputs. We input the BCJR LLRs to a CNN noise predictor to predict the signal dependent media noise. Two different CNN interfaces with the channel decoder are evaluated for TDMR. Then a second pass of the BCJR is provided with the estimated media noise and feeds its output to the LDPC decoder. The system exchanges LLRs between BCJR, LDPC, and CNN iteratively to achieve higher areal density (AD). The simulation results are performed on a GFP model with 11.4 Teragrains per square inch (Tg/in^2). For the GFP data with 18 nm track pitch (TP) and 11 nm bit length (BL), the proposed method for TDMR achieves 27.78% AD gain over the 1D PDNP. The presented BCJR-LDPC-CNN turbo detection system obtains 3.877 Terabits per square inch (Tb/in^2) AD for 11.4 Tg/in^2 GFP model data which is among the highest ADs reported to date. The second part (Chapter 3) investigates DNN-based turbo-detection for multilayer magnetic recording (MLMR), an emerging HDD technology that employs vertically stacked magnetic media layers with readers above the top-most layer. The proposed system employs two layers with two upper-layer tracks and one lower-layer track. The reader signals are processed by CNNs to separate the upper- and lower-layer signals and equalize them to 2D and 1D PR targets, respectively. The upper and lower layer signals feed 2D and 1D BCJR detectors, respectively. The detectors' soft outputs feed a multilayer CNN-based media noise predictor, whose predicted noise outputs are fed back to the BCJR equalizers to reduce their BERs. The BCJR equalizers also interface with LDPC decoders. Additional BER reductions are achieved by sending soft-information from the upper-layer BCJR to the lower-layer BCJR. Simulations of this turbo-detection system on a two-layer MLMR signal generated by a grain switching probabilistic (GSP) media model show density gains of 11.32% over a comparable system with no lower layer, and achieve an overall density of 2.6561 Tb/in^2. The third part (Chapter 4) considers a turbo-detection system that includes a CNN-based equalizer, a BCJR trellis detector, a CNN-based media noise predictor (MNP), and an LDPC channel decoder for TDMR in the presence of track-misregistration (TMR). The input readings are passed to a 2D PR equalizer which is either linear or CNN-based. The equalized waveforms are inputs to a 2D BCJR detector, which generates LLR outputs. The CNN MNP is provided with BCJR LLRs to estimate signal-dependent media noise samples and feed them back to the BCJR. A second pass through the BCJR produces LLRs which are decoded by an LDPC decoder; achieved AD is computed from the LDPC code rate. Spatially varying read- and write-TMR models are developed. We investigate the performance of the proposed system on simulated TDMR readback waveforms generated by GSP simulations. We have two types of GSP datasets. The dataset #1 includes two 10 nm BL datasets with 18 and 24 nm TP. The dataset #2 has 11 nm BL and 15 nm TP. The comparison baseline is a 1D BCJR detector with PDNP and soft intertrack interference (ITI) subtraction, referred to as 1D PDNP with LLR exchange. The write- and read-TMR are modeled as cross-track-independent downtrack-correlated random processes. In presence of joint write- and read-TMR, the proposed turbo-detection system achieves 8.34% and 0.70% AD gain over 1D PDNP with LLR exchange for TP 18 and 24 nm dataset #1 respectively, and is more robust to TMR compared to the baseline.

Deep Learning-based Turbo-detection and Equalization for Two- and Three-dimensional Magnetic Recording

Files and links (1)

Abstract

Metrics

Details