The anomaly detection is a specific branch on the data analysis that focuses on finding and highlighting the outliers of the dataset. It is useful across a range of real-world settings, including cyber-security, manufacturing, fraud detection, medical imaging, and other. In particular, in a multivariate time-series the anomalies are both depending on the state of a single time-series and the reciprocate distance between features. This thesis aims to find the anomalies in a dataset derived from the sysmetric, an Oracle system view, that collects database health and performances every minute. To achieve this goal we collected, cleaned and preprocessed the data in order to test different unsupervised machine learning models -abod, feature bagging, pca, gaussian mixture model, cblof, isolation forest- and compared the results with a validation dataset previously cleaned and preprocessed.