Advances in Nonlinear Computational Substructuring for Real-Time Hybrid Simulation
Loading...
Authors
Bas, Elif Ecem
Issue Date
2020
Type
Dissertation
Language
Keywords
data transfer , deep learning , machine learning , metamodeling , real-time hybrid simulation , seismic response
Alternative Title
Abstract
Hybrid Simulation (HS) is an advanced experimental method to investigate the overall structural behavior as well as individual element responses under realistic loading such as earthquake excitation. A typical HS configuration comprises two components: a physical or experimental substructure, which is used to simulate components that cannot be modelled numerically with full accuracy. The second component is the computational substructure for the parts that can be easily modeled with accuracy and simulated using a computational method such as finite element (FE). The dynamic analysis of this coupled experimental-computational model is usually solved using direct integration algorithms. Thus, both advantages of numerical modeling and experimental testing are leveraged to get more reliable, cost-effective, and realistic response of larger or complex structural systems. Although the numerical modeling in engineering applications is advancing with improvements in the fields of high-performance computing, there are still limitations when used in real-time HS (RTHS) applications related to the dynamic analysis solution. Understanding limitations of computational substructures and exploring new approaches in modeling such as machine learning (ML) can advance RTHS applications and capabilities. The overarching goal of this doctoral research is to identify the challenges of complex nonlinear computational substructures in RTHS and establish a new ML-based method that could potentially improve the applicability of RTHS.This study consist of three major parts. First, a compact HS setup was developed and assembled at the University of Nevada, Reno for tackling challenges and developing new concepts for computational substructuring in HS/RTHS. Once the system has been developed, a comprehensive pure analytical and RTHS testing campaign was conducted to identify the limitations and assess the performance of RTHS with nonlinear computational substructures that involve both stiffness and strength degradation. The observed limitations related to computational substructuring are strongly tied to real-time numerical integration, which further propagate when combined with RTHS typical hardware and communication loops. Such limitations called for the need for more robust and innovative solutions that could benefit from metamodeling concepts to bypass numerical integration, which motivated the third part of the study. Using ML algorithms to train and develop metamodels that represent the dynamic behavior of the computational substructure in the RTHS loop has been proposed for the first time in this study. Deep long-short term memory (LSTM) networks have been considered for the metamodeling concept. However, LSTM networks are better developed and used in advanced platforms such as Python as opposed to classical platforms used in RTHS such as Simulink/Matlab. Thus, one challenge was to implement Python-based LSTM models for RTHS analytical substructures. A novel communication scheme was developed to enable the use of Python as a computational driver from both a local computer within the RTHS loop and a cluster, i.e. high-performance computers. This communication was verified and validated when complex ML algorithms, e.g. LSTM, are involved in the RTHS loop. For completeness, the last part of the study involved RTHS of simple physical substructures (linear elastic steel brace) using the Python LSTM computational models. These tests aimed at assessing the quality assessment of RTHS tests and results when the new ML-based approach is used, and identify research needs for future research.