Dissertation
MACHINE LEARNING TECHNIQUES FOR TURBO EQUALIZATION OF TWO-DIMENSIONAL MAGNETIC RECORDING
Doctor of Philosophy (PhD), Washington State University
01/2020
Handle:
https://hdl.handle.net/2376/111181
Abstract
To reach higher areal density for next generation hard disk drives,
two-dimensional magnetic recording (TDMR) is a promising technology that emphasizes signal processing techniques for data recovery without requiring radical changes of the recording media. In TDMR, an array of read heads captures effects from multiple tracks. This helps mitigate inter-track interference, which becomes severe at higher recording densities, through two-dimensional signal processing. This work proposes machine learning approaches to address common challenges in TDMR.
First, a 2D decoding system that uses readback data from three tracks to detect data on the center track is developed.
This system consists of a linear minimum mean square error (MMSE) equalizer with partial response (PR) target, a Bahl–Cocke–Jelinek–Raviv (BCJR) detector, and an irregular repeat-accumulate (IRA) low-density parity-check decoder. A joint design method of PR target and MMSE filter coefficients to minimize detector bit error rate (BER) at the BCJR output is presented.
Next, a separate signal processing path consisting of a linear PR equalizer, a trained local area influence probabilistic (LAIP) detector and an IRA decoder is added to a 2D version of the above system for passing soft information to the BCJR for simultaneous three-track detection. This triples the system throughput. The LAIP training methods are tailored for PR equalizer outputs.
Furthermore, a deep neural network (DNN) based {\\em a posteriori} probability (APP) system that consists of an MMSE equalizer, a DNN detector, and an IRA decoder is presented. Three different types of DNNs are investigated -- convolutional neural networks, long short-term memory, and fully-connected neural networks. Turbo loops between the DNN detector and the IRA decoder are explored, and a novel DNN training-per-iteration approach for iterative decoding with the IRA is proposed. A 30.47% reduction in detector BER, 21.72% increase in areal density and three times throughput gain are achieved.
Lastly, an alternative system is developed in which the linear MMSE equalizer is replaced by a neural network based nonlinear equalizer adapted using cross entropy between the true probability of the bit and the detector’s estimate of it. Simulation results show that cross entropy is a superior criterion to MSE, and nonlinear equalizer structure is better than linear structure.
Metrics
16 File views/ downloads
40 Record Views
Details
- Title
- MACHINE LEARNING TECHNIQUES FOR TURBO EQUALIZATION OF TWO-DIMENSIONAL MAGNETIC RECORDING
- Creators
- Jinlu Shen
- Contributors
- Benjamin Belzer (Advisor)Krishnamoorthy Sivakumar (Advisor)Thomas Fischer (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Electrical Engineering and Computer Science, School of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Number of pages
- 154
- Identifiers
- 99900581414501842
- Language
- English
- Resource Type
- Dissertation