The main purpose of Structural Health Monitoring (SHM) is the assessment of structural conditions in aerospace, mechanical and civil systems. In structural engineering, damage is defined as any permanent change in the structural and geometric properties of a system caused by an external action. Vibration-based damage assessment methods rely on the use of sensors that record the structural dynamic response of a system that is determined by its structural and geometric properties. External disturbances and environmental conditions in which the system operates cause fluctuations of these properties and might hide the change in signature induced by damage. To handle the uncertainties in the determination of the structure’s characteristics, a statistical pattern recognition approach is presented in this thesis. Any statistical approach relies on the statistics of some features that provide a compact representation of the structural properties and that are sensitive to damage. Such features are called damage sensitive features and are extracted from the dynamic response of the structure: their statistical distribution is then analyzed to assess the occurrence of damage. This dissertation focuses on the analysis of the statistical distribution of damage sensitive features which are extracted through parametric and nonparametric algorithms. Cepstral coefficients are features defined in the field of acoustics and, in this thesis, they have been adapted to SHM analyses in order to develop compact damage sensitive features whose extraction requires a low computational effort. In this thesis, cepstral coefficients have been mathematically transformed through a Principal Component Analysis in order to generate damage sensitive features that are barely sensitive to measurement noise, environmental conditions and different excitation sources. In an attempt to develop an automated strategy for structural damage assessment, the search for damage sensitive features has been extended to the estimation of structural mode characteristics obtained through an output-only version of the Inner Product Vector methodology, e.g. considering only the structural response time histories. This new damage assessment procedure requires low computational effort and is capable to identify both the presence of damage and its location. However, one of the critical points of the proposed procedure consists in the manual evaluation of the spectral content of the dynamic responses that requires the user’s intervention. To automatize this procedure, a Bayesian clustering algorithm and a classifier have been successfully implemented and tested. Finally, the robustness of Bayesian regression algorithms to overfitting led us to consider their applicability to the field of system identification in order to provide a reliable estimate of the structural modal parameters that can be used as damage sensitive fea- tures. In fact, one of the main problems of system identification algorithms is that they rely on a regression algorithm that tends to overfit data producing unreliable results. Results provided by the Bayesian regression based system identification algorithm are obtained and compared with the ones coming from standard system identification algorithms.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-n50y-e138 |
| Date | January 2020 |
| Creators | Morgantini, Marcello |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0018 seconds