System identification methods are especially attractive to structural engineering for the identification of vibration modes and mode shapes of structures, as well as for detecting changes in their vibration characteristics, both under real operation conditions. Focusing on the class of subspace-based methods, the goal of this thesis is to improve the efficiency and robustness of system identification and fault detection of large in-operational structures, which are heavily instrumented and work under noisy and varying environmental conditions. In this thesis, four different hurdles are cleared. Firstly, an algorithm is derived for directly merging sensor data from multiple measurement sessions at different sensor positions and under different excitation conditions. With a modular and memory efficient approach, global subspace-based system identification of large structures is possible. Secondly, a reformulation of a least squares problem leads to a significantly faster computation of system identification results at multiple model orders, which is used to distinguish true physical modes of a structure from spurious modes under in-operation conditions. Thirdly, a statistical subspace-based fault detection method is improved using a residual that is robust to changes in the unmeasured ambient excitation. Finally, a statistical damage localization test is derived, where required sensitivities are computed from measured data without the need of finite element model. The proposed methods are validated on simulations and are successfully applied to system identification and damage detection of several large-scale civil structures.
