Deep learning for seismic structural monitoring by accounting for mechanics-based model uncertainty

This paper presents a hybrid deep learning methodology for seismic structural monitoring, damage detection, and localization of instrumented buildings. The proposed methodology develops mechanics-based structural models to generate sample response datasets by accounting for the uncertainty of model parameters that can highly affect the estimation of baseline model nonlinear responses. The uncertainty of model parameters is evaluated through the design of experiments methodology by employing the central composite design for sampling. The generated sample response dataset is utilized for training a hybrid data-driven model that combines a convolutional neural network and wavelet packet transform modules for feature extraction. The global story-level noise-contaminated response measurements are used as input for the datadriven model to perform damage detection and localization in a manner consistent with performance-based design criteria. The performance of the proposed methodology is studied in the context of numerical and experimental case studies developed based on the shake table testing of a concentrically braced frame subject to various input ground motion intensities at the E-Defense facility in Miki, Japan.