Predictive Maintanence for Datacenters

Machine learning and artificial intelligence are evolving rapidly nowadays, and application of these techniques gives a lot of benefits to companies. This thesis researches opportunity of upgrading efficiency of existing systems which implements the role of monitoring for datacenters, by adding the function of prediction the failures which can occur in period of time. In this research we applied the possibility of classification the failure of Hard Disk Drives in servers, because it is one of the main components in server infrastructure and losing data can lead to financial loses and business outages. As a metrics for monitoring were used the parameters of self-monitoring systems in Hard Disk Drives, these metrics were used as input into prediction model. As part of this thesis we reviewed different classification algorithms and chose the algorithm which fits best to our case, techniques of working with imbalanced datasets, data cleaning techniques and covered the concepts of neural networks. We implemented the random forest classifier for predicting the probability of failure. Our tests showed 80% accuracy, that means that our model can be applied in production. The functionality of the model can be extended and used for wider area, for example we can choose the metrics for servers productivity, collect the data about servers failure and train a model on these parameters and after that we can predict the failure of servers.