Real-Time Hybrid Simulation with Deep Learning Computational Substructures: System Validation Using Linear Specimens

Elif Ecem Bas,Mohamed A Moustafa

doi:10.3390/make2040026

Elif Ecem Bas, Mohamed A Moustafa

Open Access

https://doi.org/10.3390/make2040026

Copy DOI

Abstract

Hybrid simulation (HS) is an advanced simulation method that couples experimental testing and analytical modeling to better understand structural systems and individual components’ behavior under extreme events such as earthquakes. Conducting HS and real-time HS (RTHS) can be challenging with complex analytical substructures due to the nature of direct integration algorithms when the finite element method is employed. Thus, alternative methods such as machine learning (ML) models could help tackle these difficulties. This study aims to investigate the quality of the RTHS tests when a deep learning algorithm is used as a metamodel to represent the dynamic behavior of a nonlinear analytical substructure. The compact HS laboratory at the University of Nevada, Reno was utilized to conduct exclusive RTHS tests. Simulating a braced frame structure, the RTHS tests combined, for the first time, linear brace model specimens (physical substructure) along with nonlinear ML models for the frame (analytical substructure). Deep long short-term memory (Deep-LSTM) networks were employed and trained to develop the metamodels of the analytical substructure using the Python environment. The training dataset was obtained from pure analytical finite element simulations for the complete structure under earthquake excitation. The RTHS evaluations were first conducted for virtual RTHS tests, where substructuring was sought between the LSTM metamodel and virtual experimental substructure. To validate the proposed RTHS testing methodology and full system, several actual RTHS tests were conducted. The results from ML-based RTHS were evaluated for different ML models and compared against results from conventional RTHS with finite element models. The paper demonstrates the potential of conducting successful experimental RTHS using Deep-LSTM models, which could open the door for unparalleled new opportunities in structural systems design and assessment.

Highlights

Hybrid simulation (HS) is a well-established structural testing method that combines experimental components and analytical models simultaneously to evaluate structural elements and overall system behavior under realistic dynamic loading conditions usually from extreme events such as earthquake, wind, etc
The test results are compared in this paper with virtual real-time HS (RTHS) predictions, where metamodel was coupled with analytical experimental substructure, and pure analytical finite element (FE) solutions to access the quality of the RTHS tests when advanced metamodels are used as computational substructures
OpenFresco [32] is used as a middleware for the HS tests when analytical substructures are modeled in: (1) specialized FE platforms, e.g., OpenSees that can be run on the Host PC; and (2) Python-based

Summary

Introduction

Hybrid simulation (HS) is a well-established structural testing method that combines experimental components and analytical models simultaneously to evaluate structural elements and overall system behavior under realistic dynamic loading conditions usually from extreme events such as earthquake, wind, etc. In HS one or more numerically simulated structural components are replaced by experimental components In such case, no information on the stiffness of the experimental substructure is needed and a resisting force is fed directly to the hybrid model at each time step to solve the equation and obtain a new input for time step. The study concluded that the current integration algorithms might have limitations on conducting RTHS tests when some types of nonlinear behaviors are involved, and experiments become even more sensitive to hardware capabilities. To extend the introduced concept of using ML for RTHS and leverage enhanced ML algorithms, the authors conducted some foundational and preliminary work [29] to develop, validate, and verify the communication between Python-based deep-LSTM metamodels, used for RTHS computational substructures, and typical hardware and other RTHS system components. The test results are compared in this paper with virtual RTHS predictions, where metamodel was coupled with analytical experimental substructure, and pure analytical FE solutions to access the quality of the RTHS tests when advanced metamodels are used as computational substructures

Motivation

Modeling Assumptions

Model Parameters and Training Dataset