Predictive maintenance : predict upcoming failures via machine learning

Predictive maintenance is an increasingly important research area in accordance with the emergence of Industry 4.0. Since Hard Drive Disks (HDD) are among the most commonly replaced hardware components in today’s IT environments, their replacements are perfect to illustrate the efficiency of such predictive tools for maintenance. Although this subject is already well-referenced and many studies can be found in the literature, most of them propose a machine learning approach where the decision to replace a disk is made by machine learning methods. In this thesis, we propose to follow the two-level framework developed by Valentin Hamaide and his team, where the decision-making is handled separately from the machine learning methods. For this approach, in the first level, we build a health indicator which is a value that is mapped from pertinent SMART features and represents the health status of a disk. Depending on this value, an alarm would be triggered in the decision-making part of the second level. If we restrict our predictor to the work of Valentin Hamaide and his team, when we pull the trigger, it is a definitive decision. In this document, we propose to add a probability measure. This measure allows limiting prediction failures when the health indicator exceeds the threshold at one moment but shows better values the time after. Furthermore, the probability measure highlight failing disks which have recurrent small health variations. Using this two-level approach and the new probability based decision allows us to build a predictor with a F1_score of 0.5965 for a multi-class SVM classification method. This predictor performs well compared to literature and this score is improved by 0.0202 due to the use of probability. The data set used for this thesis come from Backblaze and contains 7058 runs of 91 days.