即時複合實驗為近十幾年來地震工程所發展出之一種新穎實驗方法，結合了數值分析模擬與實驗測試，可節省進行結構實驗之製作成本費用。憑著先進學者們在此技術的辛苦研究與推廣，即時複合實驗已漸漸發展成熟，而近年即時複合實驗之應用除房屋結構外，更拓展至橋梁、大地、水利工程等專業領域，其應用範圍已十分廣泛。由於即時複合實驗之原理為將完整結構拆分為數值子結構與實驗子結構，其中數值子結構最常見之模擬方法為有限元素分析法，有限元素分析採用逐步積分求解，須於有限的時間步長內完成計算，並將計算結果傳遞至實驗子結構進行加載，當數值模型為高度非線性或多自由度之複雜結構時，欲在有限時間內完成數值分析計算須仰賴高計算性能之硬體設備，因此以非線性複雜結構作為數值子結構成為即時複合實驗技術之一大挑戰。
有鑑於此，本研究以人工智慧的神經網路模型預測非線性結構受震反應，並進行深入研究與測試，再基於即時複合實驗原理架構，建立一種以長短期記憶(Long Short-Term Memory, LSTM)神經網路為基礎，並符合複合實驗原理架構與物理意義之神經網路輸入輸出關係，在Python環境下進行神經網路訓練並以此訓練模型取代傳統的有限元素數值模型。最終，以三層樓單跨裝置黏滯阻尼器之非線性鋼結構作為數值子結構，以Simulink建立之線性黏滯阻尼器作為虛擬的實驗子結構，在結構實驗室進行即時系統下的致動器連線測試， 且Python訓練之神經網路模型經轉檔並在Simulink Real-Time的環境下可以40 Hz之計算頻率順利地完成即時複合實驗。本研究結果顯示，機器學習所訓練的數值模型可正確地預測結構受震之非線性行為，並可以做為即時複合實驗之數值子結構，證明其應用於即時複合實驗方法之可行性，未來可做為含高度非線性結構之即時複合實驗技術發展的一種嶄新方法。

In the past decade, an innovative experimental method named real-time hybrid simulation (RTHS) has been developed and applied to earthquake engineering studies. RTHS which combines numerical simulation (numerical substructure) with physical testing (experimental substructure) has been adopted for verification of novel structural systems recently. It has been applied to studies of isolated buildings, bridges, soil-structure interaction, wind turbines, and etc. Generally, the numerical substructure in RTHS is modeled by using finite element method (FEM) which solves the equation of motion by a step-by-step integration algorithm. For RTHS, the solution of equation of motion must be solved within an extremely limited time step. Moreover, the response at the interfacial boundary between the numerical and experimental substructures must be sent to drive the experimental substructure and the force measured from the experimental substructure is sent back to the FEM model for calculating the response for the next time step. However, if the numerical model is a highly nonlinear model or a complex structure with large degrees of freedom, completing numerical analysis and computations in a given limited time step becomes extremely demanding and requires high performance computing machines. Therefore, using nonlinear complex structures as numerical substructure has become one of the critical challenges in RTHS.
In this study, modern artificial neural network with Long Short-Term Memory (LSTM) is adopted to predict the nonlinear seismic response of a structure first. The structure is built by OpenSees which is a three-story one-bay steel structure with several viscous dampers. Then, an input-output relationship based on the LSTM model is trained and validated which can be straightforwardly fit into the RTHS framework. Finally, the LSTM model is applied as the numerical substructure, and one of the viscous dampers modeled by Simulink is used as the experimental substructure. RTHS with a physical servo-hydraulic actuator and virtual experimental substructure is conducted. Experimental results show that the neural network model trained with Python can successfully be converted into a Simulink-compatible block and executed in Simulink Real-Time environment for RTHS. The minimum time step for RTHS is 0.025 second. The research results validate the feasibility of RTHS with a nonlinear numerical substructure. Further studies will focus on decreasing the time step of numerical substructure and increasing the complexity of numerical substructure.

[1] Takanashi, K., Udagawa, K., Seki, M., Okada. T., Tanaka, H. (1975), “Non-linear earthquake response analysis of structures by a computer-actuator on-line system: part 1 detail of the system”, Transactions of the Architectural Institute of Japan; 229: 77-83.
[2] Takanashi, K., Ohi, K. (1983), “Earthquake response analysis of steel structures by rapid computer-actuator on-line system, (1) a progress report trial system and dynamic response of steel beams” Bull Earthquake Resistance Structure; 1983(16): 103-109.
[3] Nakashima , M., Kato, H., Takaoka, E. (1992). “Development of real-time pseudo dynamic testing”, Earthquake Engineering & Structural Dynamics; 21(1): 79–92. 21(1), 79-92.
[4] Merritt, H. E. (1967), “Hydraulic Control Systems”, Wiley, New York.
[5] Zhao, J., French, C., Shield, C., and Posbergh, T. (2003), “Considerations for the development of real-time dynamic testing using servo-hydraulic actuation”, Earthquake Engineering and Structural Dynamics; 32(11): 1773-1794.
[6] Chen, P. C. and Tsai, K. C. (2013), “Dual compensation strategy for real-time hybrid testing”, Earthquake Engineering and Structural Dynamics; 42(1): 1-23.
[7] Chen, P.C., Tsai, K.C., Lin, P.Y. (2013), “Real-time hybrid testing of a smart base isolation system”, Earthquake engineering & structural dynamics; 43(1): 139-158.
[8] Wu, B., Bao, H., Ou, J., and Tian, S. (2005), “Stability and accuracy analysis of the central difference method for real-time substructure testing”, Earthquake Engineering and Structural Dynamics; 34(7): 705-718
[9] Bonnet, P. A., Lim, C. N.,Williams, M. S., Blakeborough, A., Neild, S. A., Stoten, D. P., and Taylor, C. A. (2007), “Real-time hybrid experiments with Newmark integration, MCSmd outer-loop control and multi-tasking strategies”, Earthquake Engineering and Structural Dynamics; 36(1): 119-141.
[10] Bas, E. E., Moustafa, M. A. (2020), “Performance and Limitations of Real-Time Hybrid Simulation with Nonlinear Computational Substructures”. Experimental Techniques; 44, 715–734.
[11] Bas, E. E., Moustafa, M. A. (2020), “Real-Time Hybrid Simulation with Deep Learning Computational Substructures: System Validation Using Linear Specimens”, machine learning & knowledge extraction. 2(4), 469-489.
[12] Xie, Y., Sichani, M. E., Padgett, J. E. and DesRoches. R (2020), “The promise of implementing machine learning in earthquake engineering: A state-of-the-art review”, Earthquake Spectra; 36(4):1-33.
[13] Huang, C. C., Loh, C. H. (2001), “Nonlinear Identification of Dynamic Systems Using Neural Networks” Computer-Aided Civil and Infrastructure Engineering; 16: 28–41.
[14] Eshkevari, S. S., Tak´, M., Pakzad, S. N., Jahani, M.(2021), “DynNet: Physics-based neural architecture design for nonlinear structural response modeling and prediction”, Engineering Structures; 229 (2021): 111582.
[15] Oh, B. K., Glisic, B., Park, S. W., Park, H. S. (2020), “Neural network-based seismic response prediction model for building structures using artificial earthquakes” Journal of Sound and Vibration; 468 (2020): 115109.
[16] Li. C.X., Li, H., Chen, Xu (2020), “A framework for fast estimation of structural seismic responses using ensemble machine learning model”, Smart Structures and Systems; 28(3): 425-441.
[17] Mulder, W. D., Bethard, S., and Moens, M. F. (2015), “ A survey on the application of recurrent neural networks to statistical language modeling”, Computer Speech & Language; 30(1): 61-98.
[18] Hochreiter, S., Schmidhuber, J. (1997), “Long short-term memory”, Neural Computation; 9(8):1735-1780.
[19] Zhang, R.Y., Chen, Z., Su Chen, Zheng, J. W., Büyüköztürk, O., Sun, H. (2019), “Deep long short-term memory networks for nonlinear structural seismic response prediction”, Computers and Structures. 220(2019): 55–68.
[20] Huang, X. H., Ye, Y. M., Xiong L., Lau, R. Y. K., Jiang , N., Wang, S.K.(2016), “Time series k -means: A new k -means type smooth subspace clustering for time series data”, Information Sciences; 367–368(2016): 1–13.
[21] Li, T., Pan, Y. X., Tong, K. T. , Ventura, C. E., Silva, C. W. de (2020), “Attention-Based Sequence-to-Sequence Learning for Online Structural Response Forecasting Under Seismic Excitation”. IEEE Transactions on Systems, Man, and Cybernetics: Systems; 52(4): 2184-2200.
[22] Kim, H. S. (2020), “Development of Seismic Response Simulation Model for Building Structures with Semi-Active Control Devices Using Recurrent Neural Network”, Apply Science; 2020(10): 3915.
[23] Xu, Z. K., Chen, J., Shen, J. X., Xiang, M. J., (2021), “Recursive long short-term memory network for predicting nonlinear structural seismic response”, Engineering Structures; 250(2022): 113406.
[24] 內政部營建署，「建築物耐震設計規範及解說」，台北，2011。
[25] Chen, P. C., Lu, Y. Y., and Ting, G. C (2021), “Numerical modelling and experimental validation of a rotary fluid viscous damper”, Smart Structures and Systems. (under review)
[26] Applied Technology Council (2009), “Quantification of building seismic performance factors”, Redwood City, USA: FEMA P695.
[27] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv:1412.6980.