Systems and methods for predicting storage device failure using machine learning

Patent metadata

Field	Value
Publication number	US-11657300-B2
Application number	US-202015931573-A
Country	US
Kind code	B2
Filing date	May 13, 2020
Priority date	Feb 26, 2020
Publication date	May 23, 2023
Grant date	May 23, 2023

Abstract

A method for predicting a time-to-failure of a target storage device may include training a machine learning scheme with a time-series dataset, and applying the telemetry data from the target storage device to the machine learning scheme which may output a time-window based time-to-failure prediction. A method for training a machine learning scheme for predicting a time-to-failure of a storage device may include applying a data quality improvement framework to a time-series dataset of operational and failure data from multiple storage devices, and training the scheme with the pre-processed dataset. A method for training a machine learning scheme for predicting a time-to-failure of a storage device may include training the scheme with a first portion of a time-series dataset of operational and failure data from multiple storage devices, testing the machine learning scheme with a second portion of the time-series dataset, and evaluating the machine learning scheme.

First claim

The invention claimed is:

1. A method for predicting a time-to-failure of a target storage device, the method comprising:
training a machine learning scheme with a time-series dataset of operational and failure data from one or more storage devices;
receiving telemetry data from the target storage device, the telemetry data including data regarding a state of the target storage device; and
inputting the telemetry data to the machine learning scheme;
wherein the machine learning scheme outputs a time-window based time-to-failure prediction for the target storage device in response to the telemetry data.

2. The method of claim 1 , wherein:
the machine learning scheme outputs at least two classes; and
one or more of the at least two classes comprise time windows.

3. The method of claim 2 , wherein training the machine learning scheme comprises splitting the time-series dataset into time windows, and a class corresponds to a time window.

4. The method of claim 3 , wherein the operational and failure data are split into time windows based on a time-to-failure.

5. A method for training a machine learning scheme for predicting a time-to-failure of a storage device, the method comprising:
pre-processing a time-series dataset of operational and failure data from one or more storage devices, thereby generating a pre-processed dataset; and
training the machine learning scheme with the pre-processed dataset,
wherein the machine learning scheme generates a time-window based time-to-failure prediction.

6. The method of claim 5 , wherein:
the time-series dataset comprises one or more features associated with a storage device; and
the method further comprises ranking at least two of the features.

7. The method of claim 6 , further comprising limiting a number features included in the pre-processed dataset, thereby reducing a dimension of the pre-processed dataset.

8. The method of claim 6 , further comprising ranking the features by at least one of recursive feature elimination, correlation attribute evaluation, gain ratio attribute evaluation, or information gain attribute evaluation.

9. The method of claim 5 , further comprising removing noise from at least a portion of the time-series dataset.

10. The method of claim 5 , further comprising modifying at least a portion of the time-series dataset by at least one of data transformation, data aggregation, or data standardization.

11. The method of claim 5 , further comprising removing at least one redundant feature of the features.

12. A method for training a machine learning scheme for predicting a time-to-failure of a storage device, the method comprising:
receiving a time-series dataset of operational and failure data from one or more storage devices;
training the machine learning scheme with a first portion of the time-series dataset;
testing the machine learning scheme with a second portion of the time-series dataset to produce a result; and
evaluating the machine learning scheme based at least in part on the result.

13. The method of claim 12 , wherein evaluating the machine learning scheme comprises calculating a performance score.

14. The method of claim 13 , wherein the performance score comprises one or more of a precision component, a recall component, or an F-score component.

15. The method of claim 13 , wherein the performance score is based on two or more components, and a maximum of the two or more components is used as the performance score.

16. The method of claim 15 , wherein one or more of the two or more components are weighted.

17. The method of claim 12 , wherein the machine learning scheme is evaluated with a cost function.

18. The method of claim 17 , wherein:
the cost function comprises a user-defined cost function; and
the method further comprises minimizing the user-defined cost function.

19. The method of claim 17 , wherein the cost function is based on a confusion matrix.

20. The method of claim 19 , wherein the confusion matrix comprises one or more weight classes for rewards or penalties associated with early or late predictions, respectively.

21. The method according to claim 5 , wherein the time-series dataset of operational and failure data from one or more storage devices includes data regarding states of the one or more storage device.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Zheng Qinling
• Elyasi Nima
• Sinha Vikas
• Choi Changho

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06F11/16

  Error detection or correction of the data by redundancy in hardware · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title