A hybrid approach for prediction of long-term behavior of concrete structures

Concrete displays long-term time-dependent behavior due to its rheological properties. The prediction of long-term behavior of concrete is difficult, even under laboratory conditions, due to the stochastic nature of its rheological phenomena. In concrete structures, long-term prediction is even more challenging due to the presence of uncontrolled conditions, such as variations in temperature, humidity, and loading. Current approaches for prediction of long-term time-dependent behavior at structural scale involve computationally intensive stochastic finite-element methods in which multiple creep and shrinkage models are implemented. However, these models are often calibrated using a database of experiments that are not the most informative for a specific structure. Structural health monitoring can improve prediction accuracy by providing structure-specific in-situ measurements of strain and temperature. Strain sensors, however, measure a multitude of effects simultaneously present in the structure, making it difficult to decouple effects of interest. In this work, a hybrid method employing probabilistic neural networks and engineering code models is proposed for the prediction of long-term behavior in concrete structures. A modular architecture is employed to decouple temperature-dependent environmental strain from long-term time-dependent strain. Generalized creep and shrinkage code models are fitted to the resulting time-dependent strain component data and used for prediction. The method is applied to a concrete pedestrian bridge instrumented with several embedded strain and temperature sensors. Excellent accuracy is achieved in the prediction of structural behavior multiple years beyond the training range. This, in turn, enables the detection of unusual structural behaviors with both gradual and sudden manifestation.