CLUSTERING DAMAGE INDICES IN DATA DRIVEN STRUCTURAL HEALTH MONITORING

The importance of structural health monitoring (SHM) in enhancing the resilience, longevity, and safety of civil infrastructure is becoming more widely acknowledged. Precast reinforced concrete (PRC) and reinforced concrete (RC) bridge girders are examples of essential assets that need sophisticated monitoring systems that can track degradation under service loads and identify early damage. This thesis explores an integrated SHM methodology that combines dynamic and static based measurements with acoustic emission (AE) techniques to determine the best sensor configuration for bridge girder monitoring. In a controlled laboratory setting, seven beams: four RC and three PRC specimens were put through a series of four-point bending tests. To record structural responses during increasing load, the monitoring system used AE sensors, displacement transducers, and piezoelectric accelerometers. Before and after cracking, dynamic tests were conducted to assess changes in modal parameters such natural frequencies and damping ratios. These changes were then compared with data on AE activity and static displacement to determine any associations with structural damage. All the features extracted are called Damage Indices. This work's main contribution is the creation of an automatic process that uses unsupervised clustering of monitoring data to identify damage states. Specifically, ground-based labels were used to validate the DBSCAN algorithm and compare it to k-means clustering. Because DBSCAN can detect arbitrary formed clusters and control noise in the dataset, the results show that it performs better than k-means in consistently differentiating between undamaged, slightly damaged, and severely damaged states. Eventually, in a round robin fashion, are considered different features per time, exploiting the so-called mask matrix, to identify the more effective ones in catching the damaged state of the structural element Nevertheless all the possible combinations of feature are not investigated, the conclusion of this master thesis offer great automatic tool that is capable to analyze the effectiveness of different sensor layouts, and find the best trade off in between cost effective scenario and reliable monitoring system.