Structural Health Monitoring is a growing area of interest given the benefits obtained from its use. This area includes different tasks in the damage identification process, among them, the most important is the damage detection at an early stage which enables to increase the security in mechanisms and systems, reducing risks and avoiding accidents. As a contribution in this topic, this work presents a data-driven methodology for the detection and classification of damages by using multivariate data driven approaches and machine learning algorithms which are validated and compared by using data from real structures in order to determine its behavior. In the methodology, PCA (Principal component analysis) and some pre-processing steps are used as the mechanisms to reduce data and build the features vector with relevant information about the different states of the structures under test. This methodology is validated by using some aluminum plates which are instrumented and inspected by means of PZT transducers attached to them and working in in several actuation phases. Results show a properly damage detection and classification of different simulated and real-damages.