遞迴歸演算法用於建築結構之系統 參數識別

Recursive algorithm for system parameter identification of building structures

鍾榮家(Rong-Jia Zhong)

指導教授 ： 王安培

中原大學/工學院/土木工程研究所/碩士(2008年)

https://doi.org/10.6840/CYCU.2008.00416

摘要

建物由非單一材料構成，耐震規範和一般設計分析大部分只考量線性行為，當遇到特殊強震或屋齡已久遠之建物會有非線性之動態行為，若考量建物非線性行為更能了解結構遇強震之動態行為和其耐震能力。且因地震波具有方向性，不同震源將使建物有不同之動態行為，若能在線上即時校正系統參數便可解決此問題。因此本文期望建立一個建物非線性行為線上即時識別模式，作為監測及控制之用。
本文建立兩種識別方法：1.採用遞迴正交最小平方(ROLS)自迴歸力法(ARX) (Auto Regressive with eXogenous using Recursive Orthogonal Least Square method)，簡稱ROLS-ARX，此法改進遞迴最小平方法(Recursive least square method)需設定巨大的初始值以及求共變異數之反矩陣之缺點，減少電算誤差及計算時間。2.採用移動窗格ARX之類神經網路形式，簡稱ANN-WARX，因方法1在遺忘因子(forgetting factor) 設定上，太大無法處理時變的資料、太小則無法有效的識別系統，為了更能掌握系統的時變性，進而建立此法，在適當的時間窗格大小中識別每個窗格之系統來估測系統的變化情形，且使用類神經網路具有良好收斂性之優點。
最後實例分析三種案例：1.設計之非線性模型、2.國家地震中心之三層樓鋼構架加裝圓棒阻尼器試驗(El Centro 650 gal發生破壞)、3.台東消防局(在民國97年4月1日餘震發生破壞)。探討非線性行為頻率與阻尼比變化，以ANN-WARX法之識別度較高，亦可得知發生非線性行為之時間。在預測建物動態行為方面，互相比較之下，ANN-WARX亦有最佳的效果。

關鍵字

遞迴正交化最小平方法 ； 系統識別

並列摘要

Building structures are made of several materials, seismic design code usually considers its linear behavior. When there is a big earthquake or old buildings, the nonlinear dynamic behavior of structures may occur during the earthquake. If we concern the nonlinearity, we will know more about the behavior and seismic capacity of structures. Besides, seismic waves may come from different directions at each earthquake. It would lead to the different dynamic behaviors of structure. Correcting the system parameters on-line can solve this problem. Thus, the purpose of this research is to develop the on-line system identification which considers the nonlinear behavior of structures to monitor and control the behavior of the building structures.
This study developed two methods: 1. Auto Regressive with eXogenous using Recursive Orthogonal Least Square method, called ROLS-ARX. This method improved the shortages of Recursive least square method because it performs without setting enormous values and calculating inverse of covariance matrix. It also reduces computing error and saves time. 2. ARX of artificial neural network using moving window, called ANN-WARX. Because the setting of forgetting factor , when it is large, ROLS-ARX method can not access the data of time-variant correctly, however, when the setting of is small, ROLS-ARX method cannot identify system effectively. To identify the system in the appropriate time window can trace the variation of the system parameters, and the artificial neural network have better capability of operation.
Finally, the design of nonlinear model, experiment of steel frame with damper and Fire Department building in Taitung are taken as three examples for analyzing the frequency and damping ratio time-varying behavior in this study. According to the results, ANN-WARX method had the best effect on tracing variation of frequency and predicting dynamic behavior of structures.

並列關鍵字

system identification ； ROLS-ARX ； ANN-WARX ； Recursive Orthogonal Least Square method

參考文獻

【1】 Anil K. Chopra, Dynamics of Structures: Theory and Applications to Earthquake Engineering. Prentice Hall; 2 edition .September 11, 2000.

【2】 D.L.YU, J.B. Gomm and D. Willams, A Recursive Orthogonal Least Squares Algorithm for Training RBF Networks. Neural Processing Letters 5: 167-176, 1997.

【3】 F.Thouverz and L.Jezequel, Identification of Narmax Models on A Modal Base. Journal of Sound and Vibration189(2),193-213, 1996.

【4】 Gersch, W. and S. luo, Discrete Time Series Synthesis of Randomly Excited Structural System Response,J. Acoust. Soc. Amer., 51, 402-408, 1972.

【5】 James E. Bobrow and Walter Murray, An Algorithm for RLS Identification of Parameters that vary Quickly with time. IEEE Transaction on Automatic Control, Vol.38, No2, February 1993.