Hybrid structural analysis integrating physical model and continuous-time state-space neural network model

Abstract

The most likely scenario for civil engineering structures is that only some components or parts of a structure are complex, while the rest of the structure can be well physically modeled. In this case, utilizing powerful neural networks to model these complex components or parts only and embedding the neural network models into the structure might be a viable choice. However, few studies have considered the real-time interaction between the neural network model and another model. In this paper, a new hybrid structural modeling strategy that incorporates the neural network model is proposed. Structures installed with energy dissipation devices (EDDs) are investigated, where continuous-time state-space neural network (CSNN) models are adopted to represent EDDs and to couple with the physical model of the structure excluding EDDs through the state-space substructuring method. First, CSNN models with an identical model configuration are trained to represent different physical models of EDDs and fit the experimental results of a damper to evaluate the CSNN model at the model level. Then, to demonstrate the hybrid structural analysis method, the CSNN-based structural models of the interfloor-damped and base-isolated structures are established for seismic analyses. It is observed that CSNN-based models exhibit high prediction performance and are easy to implement. Therefore, the developed hybrid structural analysis method that adopts CSNN models for EDDs is engineering practical.