以支持向量回歸法預測平滑遲滯模型參數之研究

Hysteretic Model Parameters with Using Support Vector Regression

林子剛(Tzu-Kang Lin) ； 楊子慧(Tzu-Hui Yang) ； 張皓惇(Hao-Tun Chang) ； 王柄雄(Ping-Hsiung Wang) ； 張國鎮(Kuo-Chun Chang)

《結構工程》 36卷3期 (2021/09) Pp. 136-152

https://doi.org/10.6849/SE.202109_36(3).0007

摘要

本研究開發一套以人工智慧為基礎的模型，將其用於預測平滑遲滯模型（Smooth Hysteretic Model, SHM）的相關參數。近年來以Bouc-Wen模型為基礎的平滑遲滯模型常被用於判定損害累積與遲滯迴圈的加載路徑，該類模型包含五種主要參數， 這些參數用以描述受撓曲為主的鋼筋混凝土（Reinforced Concrete, RC）橋柱之耐震性能。其中，與時間變化相關的參數，僅能透過實驗取得，但實際上無法頻繁使用真實結構物進行試驗，進而影響平滑遲滯模型的實用性。雖然平滑遲滯模型的方便性有待商榷，但其性能表現優於其他現有的遲滯模型，因此本研究試以支持向量回歸法（support-vector regression, SVR），結合人工智慧與平滑遲滯模型的優勢，開發一套更加完善之遲滯模型，以利橋梁震損與耐震性評估。研究資料取自近年來進行實驗之九座不同RC橋柱，樣本資料共有119筆勁度衰減參數與81筆束縮參數資料。並以80%資料進行訓練，其餘20%資料用於測試。將各橋柱的主筋比、高寬比、位移及殘餘位移做為輸入參數，透過支持向量回歸法，預測出該橋柱之勁度衰減與束縮參數。該方法能在低誤差的情況下，精準預測各項時變參數。最後，將分別以支持向量回歸法預測之參數，與實驗數據識別之參數繪製之遲滯迴圈進行比較。分析結果表明，利用本研究所提出之方法，可以進行可靠的橋柱耐震性預測，無須進行繁雜的實驗流程，即可達成協助平滑遲滯模型預測橋柱之震損程度與耐震特性。

關鍵字

支持向量機 ； 支持向量回歸 ； 平滑遲滯模型 ； 勁度衰減 ； 束縮

並列摘要

This study developed artificial intelligence-based models for predicting smooth hysteretic model (SHM) parameters. Recently, an SHM based on the Bouc-Wen model was developed to determine damage accumulation and path dependence of reloading. The model comprises five main parameters that describe the seismic behavior of ductile, flexure-dominated reinforced concrete (RC) bridge columns. However, each time-variant parameter can be derived only through practical experiments and cannot be tested on actual structures; therefore, the SHM is not very practical. In this study, support-vector regression (SVR) was adopted to capitalize on the advantages of the developed SHM, which exhibits superior performance to other existing hysteresis models. Nine different RC bridge columns were tested under displacement time histories, and a total of 119 samples were acquired. Of the samples, 80% were used for training and the remaining 20% were used for testing. The longitudinal reinforcement ratio, aspect ratio, and displacement or residual displacement of individual columns were set as the inputs to the SVR models, and the pinching and stiffness degradation parameters were set as the model output. Time-variant parameters could be predicted accurately with low deviation and error percentages. Moreover, hysteresis loops were generated using the identification parameters, and the SVR prediction results were compared with experimental data. The results indicated that the seismic behavior of the RC bridge columns could be estimated with high reliability using the proposed method without the support of experimental progress and support the SHM to predict the degree of damage.

並列關鍵字

support vector regression ； smooth hysteretic model ； pinching ； stiffness degradation