US 12,535,952 B2 · Machine learning defect management in storage devices

Abstract

Methods are provided for managing defects in Hard Disk Drive (HDD) storage devices. In particular, only a portion of the cylinders of an HDD is tested. A bag of machine learning models is used to reconstruct the data for the untested cylinders. A defect file for the HDD is generated, a classifier model may be applied to the defect file, and one or more neural network models may be applied. If the defects are unsuitable for use by the models, then a scan of the entire HDD is run instead. An HDD comprises a rotating disk and a read/write head actuated above the disk surface. The disk may be formatted into concentric data tracks, with each track being divided into sectors. The tracks may be organized into zones (groups of tracks called cylinders), and the axially parallel sectors in each cylinder may be organized into wedges.

Claims23

Dependent claims

A device, comprising:

a rotating disk having a media surface;
a head actuated over the media surface;
a plurality of concentric data tracks disposed on the disk, wherein:
the tracks are grouped into cylinders,
the tracks are partitioned into radially adjacent sectors, and
radially adjacent sectors in the cylinders are grouped into wedges;
and
control circuitry comprising a Machine Learning Defect Management (MLDM) logic, wherein the MLDM logic is configured to:
select a portion of the cylinders for testing;
test the wedges in the selected cylinders for defects, and store the results in a defect file;
determine if the defect file is suitable for classification;
classify the defects for suitability for input to a machine learning model;
input the results of the detected defective wedges to a first machine learning model performing a full-track analysis;
input the results of the first machine learning model to a second machine learning model performing a partial-track analysis;
and
infer the defective or non-defective status of the wedges in the untested cylinders with the first and second machine learning models.

13 more elements

A non-transitory computer-readable storage medium configured to store instructions and data to be executed by one or more processors, where the instructions, when executed by the one or more processors individually or in combination, cause a device, the device comprising a disk with a plurality of concentric data tracks disposed on the disk, the tracks grouped into cylinders and divided into sectors of which radially adjacent sectors are grouped into wedges, to perform steps comprising:

selecting a portion of the cylinders for testing;
testing the wedges in the selected cylinders for defects and storing the results in a defect file;
determining if the defect file is suitable for classification;
classifying if the defect file is suitable for input to a machine learning model;
inputting the defect file to a first machine learning model and performing a full-track analysis;
inputting the results of the first machine learning model to a second machine learning model and performing a partial-track analysis;
and
inferring a defective or a non-defective status of the wedges in untested cylinders with the first and second machine learning models.

5 more elements

Dependencies are inferred from the claim text.

Machine learning defect management in storage devices

Life of this patent

Machine learning defect management in storage devices

Life of this patent

Abstract

Claims23