基於確定-預測性模型之遞迴式系統識別分析與結構損傷探測

Abstract

本研究結合建構在確定-預測性模型之遞迴式系統識別與狀態空間損傷定位向量法發展結構損傷探測技術，做為發展即時結構健康監測系統之基礎。確定-預測性模型之遞迴式系統識別法係根據遞迴式子空間系統識別法，結合工具變數之概念所發展。經由數值模擬分析與國家地震工程研究中心之振動台地震模擬試驗結果顯示，吾人可在輸入–輸出訊號含高斯雜訊的混合式狀態空間模型下，透過結構全域之地震反應即時更新系統參數，作為損傷探測分析之依據，多數情況下均能定位出結構之局部受損樓層，具有相當之可信度。本文驗證R4SID法對於噪訊濾波能力之強健性，並探討初始矩陣與Hankel時序矩陣內涵與排序之較佳選擇，提供未來實際應用之參考，確認其發展成即時線上分析模組的潛力。於國家地震工程研究中心完成之振動台地震模擬試驗與損傷探測分析顯示，在完全觀測條件下，無論單一或多重樓層破壞都能成功辨識出來，部分觀測條件下在單一樓層受損時或有機會成功辨識。對於既有損傷結構(不良結構)而言，R4SID法也能成功定位出新增或損傷程度加劇的樓層，同時找出原已受損之樓層。比較研究之結果顯示，R4SID之強健性優於N4SID與RSSI-COV等系統識別方法。最後，並以進行中的結構健康監測現地試驗計畫—中部科學園區HL廠在2015年0420地震之實測動態反應資料進行系統識別分析，其識別結果與該廠ETABS結構模型預估之自然頻率有高度相關性，進一步驗證R4SID應用於科技廠房的可行性。

This study integrates the Recursive Stochastic Subspace System Identification (R4SID) for Deterministic-Stochastic Models with the method of damage localization vector (DLV) under a state-space framework for structural damage detection. This in turn serves as the basis for the development of a real-time structural health monitoring system. R4SID is developed based on the theory of Recursive Subspace Identification Algorithm, along with the concepts of Instrumental Variable. Via numerical simulations as well as the shaking table tests conducted in NCREE, the proposed scheme has been proved sufficient, with the hybrid deterministic/stochastic model, in updating the system parameters from the global seismic responses contaminated with Gaussian white noise. With the identified system parameters for damage detection, the scheme is able, in most conditions, to localize the damaged stories of the building. The R4SID proves quite robust in noise filtering. The possible choices of the initial matrix for recursive analysis and arrangement of the Block Hankel matrix have been explored to provide as references for practical application and confirm its potential for real-time system identification. The damage localization analysis of the shaking table tests conducted in NCREE indicates that, under a full observation condition, the damaged stories can be identified, regardless of single or multiple damages. Whereas only single damage could be identified under a partial observation condition. For ill - conditioned structures that have been previously damaged, the R4SID method can successfully localize new or deteriorated damages while identifying the existing damages as well. A comparative study indicates better robustness of the R4SID than both the N4SID and RSSI-COV methods. Finally, system identification of the HL fab in Taichung Science Park for an on-going site demo project on SHM using the seismic response data recorded on April 20, 2015 has been conducted. Encouragingly, the identified natural frequencies of the predominant modes are closely correlated with those obtained from the ETABS model of the fab. The feasibility of R4SID algorithm for the SHM of high-tech fabrication factories is further confirmed.