• classification
• damage localization
• damage quantification
• deep learning
• machine learning
• regression
• structural health monitoring

The Structural Health Monitoring (SHM) through the use of data collected by sensors installed on a civil structures is an increasingly central topic. SHM aims to detect breaks and measure the degree of damage, in order to act in time with maintenance interventions.
The research done in this area increasingly aims to introduce Machine Learning (ML) techniques to perform this task.
Adopting a Supervised Learning (SL) approach allows to make a more accurate data analysis, in order to better understand the level of damage with respect to an Unsupervised Learning (UL) approach. The possibility of carrying out more in-depth analysis using a supervised approach has an important disadvantage: not having the Ground Thruth knowledge, that means not having available labeled data.
In this thesis work, after a description of the most adopted methods based on an UL approach and their limitations, Some Deep Learning (DL)-based techniques are leveraged to solve a Simulation-Based Classification (SBC) and Simulation-Based Regression (SBR) problems. A synthetic dataset was generated by exploiting some parametric Model Order Reduction (MOR) techniques, thanks to which we can simulate the behavior of the dynamic response of the sensors to the solicitations starting from a model of the structure. After proposing some DL architectures for SBC and SBR problems, an evaluation on their performance is done to study how they work on artificially corrupted data, which simulate some real distortions.