SEISMIC RESPONSE PREDICTION OF BUILDINGS USING AUGMENTED KALMAN FILTER AND DEEP LEARNING
Document Type
Book
Publication Date
2024
Abstract
Real-time seismic response prediction methods could be computationally expensive and inadequate, therefore precise, and faster prediction methods are required. Responses of a structure are necessary for rapid estimation of human injury, occupancy level of buildings, and property loss. Over the past few decades, big data has been collected by acceleration sensors of seismic activity during strong ground motions. This historically recorded data could be sufficient to be coupled with the latest artificial intelligence methods to provide accurate surrogate models that can predict seismic responses in real-time and provide engineering judgment on the response level of a building at different floors. Physics-based relationships do not typically capture all important phenomena of complex systems and could be regarded as incomplete representations. Some parameters are already not included, which could eventually affect the behaviour of model output, or there is a calibration process that is required such that the physics-based relationship satisfies unique cases (e.g., stiffness, and damping). Therefore, it is worthwhile to discover advances in hybrid empirical methods that more effectively interpret field data. Kalman filtering is a useful structural health monitoring (SHM) algorithm that uses sensor measurements over time to produce estimates of unknown states of a system. Since the Kalman filter is famous for its efficiency and reliability in linear system estimation with noisy measured data, the augmented Kalman filter (AKF) is considered here for an observed building (since measurements are available). Since a deep learning architecture (ConvLSTM) can effectively predict the responses of observed floors, it will be compared with AKF. The primary difference between AKF and ConvLSTM during deployment is that the former requires ground and top-floor observations to estimate states of unobserved stories, whereas the latter only requires ground acceleration observations. The following is deduced from this study: 1) MDOF analysis requires calibrations for stiffness and equivalent damping, however, calibrations are not required with the ConvLSTM model; 2) Kalman filter methods aided by sensors could provide accurate system identification, however, ConvLSTM can perform synthetic simulations responses without observations from superstructure sensors (only input ground motions).
Recommended Citation
Torky, Ahmed A. PhD and OHNO, Susumu Ph.D., "SEISMIC RESPONSE PREDICTION OF BUILDINGS USING AUGMENTED KALMAN FILTER AND DEEP LEARNING" (2024). Civil Engineering. 246.
https://buescholar.bue.edu.eg/civil_eng/246