Evaluating One-Class Classifiers for Fault Detection in Hard Disk Drives

• Main Authors: Pereira, Francisco Lucas F., Teixeira, Daniel N., Gomes, Joao Paulo P., Machado, Javam C.
• Format: Conference Proceeding
• Language: English
• Subjects: Data models; Fault detection; Hard Disk Drives; Hard disks; Neural networks; one-class classifiers; Reconstruction algorithms; Support vector machines; Training data
• ISSN: 2643-6264
• DOI: 10.1109/BRACIS.2019.00108

In recent years, both academy and industry have focused on designing solutions for fault detection in Hard Disk Drives (HDD). Such effort is justified by the wide range of benefits provided by knowing, in advance, when a failure will occur. By being able to predict failures on HDD, one can avoid data losses and also improve maintenance planning. Recently, several works achieved remarkable performances by using SMART (Self-Monitoring Analysis and Reporting Technology) attributes along with standard classification or regression algorithms. However, its is well known that data from degraded HDD is scarse, and because of that, using such methods may be unfeasible in practical applications. In this regard, one class classifiers arise as a good alternative since it can be used to detect anomalous instances by modeling the behavior of a healthy HDD. In this paper, we aim to evaluate common one-class classifiers for fault prediction in HDDs. In our experiments, we consider density, boundary and reconstruction methods. Based on our results, we can verify that reconstruction methods seem to be more adequate for the problem under analysis.