Health monitoring of dynamic systems via data-driven approaches

A worldwide push for sustainability, advances in technology and changes in infrastructure demand and usage, combined with more frequent and intense extreme events, are redefining design and monitoring practices across civil, mechanical and aerospace systems. For civil infrastructure systems, this shift highlights a national need to develop monitoring and diagnostic systems able to identify damage at its early stages, preventing it from evolving into catastrophic failures and ensuring the safety of the monitored structures. Additionally, early damage detection enables the optimization of maintenance activities and extends the system's service life by allowing for the planning of targeted interventions, thereby minimizing costs and service downtime. Within this context lies Structural Health Monitoring (SHM), which encompasses the activities and methodologies necessary for assessing structural integrity. SHM applies across various sectors, including aerospace, civil and energy, where assets, even those with similar dynamic traits, can exhibit significant variability. Importantly, such variability also exists within a single field, motivating the development of monitoring strategies that are independent of any specific structure. To address this need, the present work proposes a set of data-driven algorithms for damage identification. These methods analyze the evolution of the structural state by comparing current data to a reference condition, built from the historical data collected by sensors installed on the monitored structure, without requiring a prior physical model. Moreover, the classification of structural conditions is carried out in an unsupervised fashion, resulting in high flexibility as it can detect previously unknown structural variations. Specifically, two damage identification algorithms are developed in this work, both within a vibration-based SHM framework and grounded in advanced signal processing. These algorithms rely on the analysis of the structure’s dynamic response with the aim of identifying and extracting Damage Sensitive Features (DSFs), parameters that are expected to be sensitive to damage while remaining robust to variations in loading and environmental conditions under which the monitored system operates. Both methodologies have been designed to minimize user intervention and to automate the monitoring process. Furthermore, their complementary performance under different types of structure and loading conditions improves the detection of various damage scenarios and ensures greater flexibility of the monitoring system. The first proposed algorithm is based on the sequential application of principal component analysis and independent component analysis, aiming to isolate the effects introduced by damage in the structural response from those related to variations in environmental and loading conditions. The second algorithm uses discrete wavelet analysis to extract DSFs that are sensitive to both distributed and localized structural changes. The results obtained for the inspected bridge test cases demonstrate that these data-driven techniques provide generalizability across various sectors, effectively filtering out the effects due to the load. To evaluate the applicability of the proposed methodologies to mechanical structures, their performance has been tested on experimental data from a scaled model of a wind turbine blade.