• Citer

Model Order Reduction methods for sensor data assimilation to support the monitoring of embankment dams

par Nasika, Christina
Président du jury Berke, Peter
Promoteur Gerard, Pierre ;Massart, Thierry,Jacques
Co-Promoteur Díez, Pedro;Zlotnik, Sergio
Publication Non publié, 2022-09-15

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

The latest monitoring and asset management technologies for large infrastruc- tures involve digital representations that integrate information and physical models, exist in parallel to the real-life structures, and are continuously up- dated based on assimilated sensor data, in order to accurately represent the actual conditions in the structures. This type of technology is often referred to as Digital Twin. The implementation of such cutting-edge technology in monitoring assets like tailings dams, or embankment dams in general, and other large structures, implies the development of highly efficient numerical tools that, combined with sensor data, may support rapid, informed decision making.For the particular case of embankment dams, enabling this type of tech- nology requires an efficient numerical model that describes the coupled hydro-mechanical phenomena, pertinent to a dam functioning and safety. This may for instance be a Finite Elements (FE) model, describing the groundwater flow through unsaturated porous geomaterials.The process of updating and calibrating a model, such as the above mentioned FE model, based on sensor data is typically referred to as data assimilation. Often, this is achieved via an optimization approach, where a specific problem is solved multiple times for various parametric values, in search for the values that best describe the sensor data. The bottleneck in this type of application is typically the cost of multiple evaluations of the model, that may become prohibitive when the underlying FE model is large. In order to enable such applications, the present work proposes Model Order Reduction (MOR) methods tailored to the hydro-mechanical nonlinear problem at hand.MOR aims at the creation of a surrogate model that seeks an approximation of the FE solution in a reduced-order space. This is achieved by applying an offline-online strategy. In the offline stage, the solution manifold of thefull-order problem is sampled, in order to identify a low-order affine subspace, where an accurate approximation of the full-order solution can be captured. To tackle the nonlinearities related to partially saturated conditions in the soil, a similar strategy must be employed in order to define reduced-order spaces where an affine system approximation may be recovered. The resulting Reduced Order Model (ROM) may be used as an efficient surrogate to the FE model in any problem that requires fast and/or repetitive solutions.In this work, MOR techniques are implemented to solve the coupled nonlinear transient problem under consideration. ROMs are created to solve problems that pertain to tailings dams and embankment dams monitoring. The efficiency and the accuracy of these models are demonstrated by solving inverse problems for parametric identification. MOR is found to be a reliable tool, significantly accelerating the inverse identification process while resulting to accurate solutions.