Unsupervised Anomaly Detection for Hard Drives

In the age of smartsensors and industry 4.0 continuous monitoring of different machinery produce enormousamount of data, because of that datacenters are now-a-days a very important asset not only for large scale cloud providers, but also for medium to large enterprises,which decide to store in-house the ever increasing data collected during business operations. An efficient method for the maintenance of the great number of hard-drives housed in datacenters is critical to assure avaiability in a cost effective manner.Since 2013, Backblaze \url{https://www.backblaze.com/} has published statisticsand datasets for researchers to gain insights on hard drive performaces andtheir failures, in this paper more than 2.5 million records, following hard-drives S.M.A.R.T readings for over a year, will be analyzed. The objective of this paper is to show that it is possible to build a completelyunsupervised pipeline which produces an anomaly score that highlycorrelates to hard drives time to failure (TTF), in such a way a decisionto replace them can be made before failure, with minimal waste due to false alarms. Favorable comparisons with state of the art supervised classifiers will be presented. A brief example of how such a pipeline can beextended for data streams and continuos sensor monitoring will be given.