Communication Development and Verification for Python-Based Machine Learning Models for Real-Time Hybrid Simulation

Abstract

Hybrid simulation (HS) combines analytical modeling with experimental testing to provide a better understanding of both structural elements and entire systems while keeping cost-effective solutions. However, extending real-time HS (RTHS) to bigger problems becomes challenging when the analytical models get more complex. On the other hand, using machine learning (ML) techniques in solving engineering problems across different disciplines keeps evolving and likewise is a promising resource for structural engineering. The main goal of this study is to explore the validity of ML models for conducting RTHS and specifically introduce and validate the necessary communication schemes to achieve this goal. A preliminary study with a simplified linear regression ML model that can be readily implemented in Simulink is presented first to introduce the idea of using metamodels as analytical substructures. However, for ML, commonly used platforms for RTHS such as Simulink and MATLAB have limited capacity when compared to Python for instance. Thus, the main focus of this study was to introduce Pythonbased advanced ML models for RTHS analytical substructures. Deep long short-term memory networks in Python were considered for advanced metamodeling for RTHS tests. The performance of Python can be enhanced by running the models using highperformance computers, which was also considered in this study. Several RTHS tests were successfully conducted at the University of Nevada, Reno, with Python-based ML algorithms that were run from both local PC and a cluster. The tests were validated through comparisons with the pure analytical solutions obtained from finite element models. The study also explored the idea of embedding the delay compensators within the ML model for RTHS.

State of the art and development trends in numerical simulation for real‐time hybrid simulation

“…Using deep learning, it is possible to simulate the dynamic properties of linear and nonlinear structures and describe the structure as a black box for numerical substructure simulations. Bas and Moustafa 14,47 first applied deep learning to RTHS, using the Python-based deep long short-term memory (LSTM) method for the simulation of nonlinear numerical substructures to build a deep learning-based RTHS architecture. The training data set was obtained from OpenSees simulations of the structure under seismic excitation.…”

“…The dynamic properties of the specimens are obtained by loading the physical specimens into the data set. The use of the deep learning method 13,14 to model the input excitation and dynamic response of numerical models is an innovative approach. Compared with traditional FEM modeling methods, deep learning modeling can characterize numerical substructures more easily.…”

Implementation of offline iterative hybrid simulation based on neural networks

“…Artificial NNs provide a way to improve RTHS methods. Bas and Moustafa 20,21 Chen et al 22 built a deep learning-based RTHS architecture, which used the LSTM network model as a numerical substructure. Tang et al 23 repeatedly loaded the physical substructure and established a constitutive agent model of its reaction based on the NARX network.…”