A new physics-guided data assimilation framework for online structural monitoring: application to shaking table tests

The modified Constitutive Relation Error (mCRE) is a model updating functional whose physics-based construction avoids the explicit dependency into a priori user’s expertise. Its robustness to measurement noise and remarkable convexity properties make it a credible alternative to classical model updating methods. In recent works, the authors developed an automated mCRE-based model updating framework for the offline update of finite element models of structures submitted to ground motion inputs, before deriving it to data assimilation with the development of the Modified Dual Kalman Filter (MDKF). In the latter, we addressed sequential data assimilation by integrating the mCRE as new observer within a dual Kalman filter. Contrary to classical Kalman filters, the comparison between measurements and model predictions is achieved through the mCRE functional itself, in which the data-to-model distance is enriched with a model error term with strong mechanical content (the CRE). In this contribution, we focus on MDKF numerical improvements to achieve real-time data assimilation for structural health monitoring applications where an extensive number of parameters may need to be tracked with a few amount of data, which makes the identification process ill-posed and prone to instabilities. Using (i) parallelized mCRE computation and sigma-points propagation, (ii) the Scaled Spherical Simplex Kalman filter structure and (iii) a CRE-based clustering pre-step, CPU requirements are reduced by partial state update without loss of accuracy. The MDKF methodology and the novel key ingredients for efficient (near-optimal) data assimilation are illustrated throughout (simulated) earthquake engineering examples of growing complexity.