High Fidelity Probabilistic Structural Health Monitoring

In the last decade, sensor based Structural Health Monitoring has become an important area of research as it shows great potential for life-safety and economic benefits for improved and responsible management of our aging civil infrastructure. Structural health monitoring involves the detection of damage or deterioration within a structure, the identification of its location and severity and ultimately it can assist in providing a prognosis for the future life of a structure. Many important civil structural systems are large and complex, with many joints and components. The task of identifying damage and deterioration in a high fidelity model of such a large system based on relatively few sensor measurements which themselves are not perfect is highly challenging. The main goal of this research project is to provide more detailed and accurate estimates of structural condition in a probabilistic format, which can be integrated into more reliable decision-making tools for infrastructure stakeholders. Furthermore, development of novel algorithmic tools for nonlinear, high dimensional system identification and parameter learning are of utmost interest in many fields of engineering, biology and even finance. Thus, this research could not only benefit from knowledge in other fields, but could contribute to research domains well beyond the structural monitoring context at the heart of the motivation here.

Within Bayesian estimation, the core framework adopted in this research, the algorithmic frontiers lie in the sensitivity to noise, Gaussian versus non-Gaussian, for example, as well as tackling high dimensional problems with local nonlinearities and many static parameters to be identified. Overcoming these challenges requires both a deep understanding of well accepted algorithms as well as development of enhanced or novel algorithmic tools. In Bayesian estimation, two main filtering algorithms are extensively studied in the literature. The consequences of the Gaussianity assumption used in the unscented Kalman filter will be carefully studied on systems of interest, while enhancements of the particle filter, which behaves poorly in high dimensional systems, will be introduced. Strategies to tackle the high dimensionality issue are based on the Rao-Blackwellisation principle as well as partitioning schemes. Off-line algorithms could also be integrated in this framework to obtain more accurate estimates with lower uncertainties. Finally these schemes should be validated on realistic, experimental data.