Convolutional Neural Network Based Symbol Detector for Two-Dimensional Magnetic Recording

Conventional detection systems in hard disk drives (HDD) typically include a 2D partial response (PR) equalizer that pre-processes the readback signals and shapes the output to a controlled target response, followed by a maximum likelihood (ML) or maximum a posteriori (MAP) detector which outputs log-likelihood ratios (LLRs) to be passed to a channel decoder. Pattern dependent noise prediction (PDNP) algorithm [1] is usually incorporated into the metric computation of the trellis in the ML/MAP detector to combat media noise intrinsic to the magnetic recording (MR) channel. For next generation two-dimensional magnetic recording (TDMR) HDDs, such conventional systems would suffer from impractically large trellis state cardinality when performing multi-track detection, and they may no longer be capable of handling the increased nonlinearities in high density recording channels. This work investigates applying advanced machine learning techniques to TDMR. Convolutional neural networks (ConvNets) are employed in place of the PR equalizer and ML/MAP detector with PDNP to directly process the un-equalized readback signals and output soft estimates. ConvNets are special deep neural networks (DNNs) that assume the inputs are images and perform convolution instead of affine function in the network forward pass [2]. This enables far fewer parameters in ConvNets than regular DNNs of the same depth and therefore allows for deeper networks. The motivation to use ConvNets is the resemblance between data detection problem in MR and typical image processing problems. In MR channels, the write process converts temporal data into spatial patterns recorded on a magnetic medium, which transforms sequential correlation into spatial ISI/ITI. Data detection can be viewed as an image processing problem, proceeding from the 2D image of the shingled bits (see Fig. 1), to higher level abstractions of features by means of convolutional layers that finally allow classification of individual bits. Several variations of ConvNets are compared in terms of network complexity and performance. The best performing ConvNet detector can provide data storage density of up to 3.7489 Terabits/in
²
on low track pitch TDMR channel simulated with a grain flipping probabilistic (GFP) model.