Department of Architecture and Civil Engineering, Toyohashi University of Technology, Toyohashi 441-8580, Japan.
Sensors (Basel). 2023 May 12;23(10):4694. doi: 10.3390/s23104694.
This study proposes a methodology to predict the damage condition of Reinforced Concrete (RC) resisting-moment frame buildings using Machine Learning (ML) methods. Structural members of six hundred RC buildings with varying stories and spans in X and Y directions were designed using the virtual work method. Sixty thousand time-history analyses using ten spectrum-matched earthquake records and ten scaling factors were carried out to cover the structures' elastic and inelastic behavior. The buildings and earthquake records were split randomly into training data and testing data to predict the damage condition of new ones. In order to reduce bias, the random selection of buildings and earthquake records was carried out several times, and the mean and standard deviation of the accuracy were obtained. Moreover, 27 Intensity Measures (IM) based on acceleration, velocity, or displacement from the ground and roof sensor responses were used to capture the building's behavior features. The ML methods used IMs, the number of stories, and the number of spans in X and Y directions as input data and the maximum inter-story drift ratio as output data. Finally, seven Machine Learning (ML) methods were trained to predict the damage condition of buildings, finding the best set of training buildings, IMs, and ML methods for the highest prediction accuracy.
本研究提出了一种使用机器学习 (ML) 方法预测钢筋混凝土 (RC) 抗弯矩框架建筑损伤状况的方法。使用虚功法设计了 600 栋具有不同楼层和 X、Y 方向跨度的 RC 建筑的结构构件。使用 10 个谱匹配地震记录和 10 个缩放因子进行了 6 万个时程分析，以涵盖结构的弹性和非弹性行为。建筑物和地震记录被随机分为训练数据和测试数据，以预测新建筑物的损伤状况。为了减少偏差，多次进行建筑物和地震记录的随机选择，并获得准确性的平均值和标准差。此外，使用基于地面和屋顶传感器响应的加速度、速度或位移的 27 个强度指标 (IM) 来捕获建筑物的行为特征。ML 方法使用 IMs、楼层数以及 X 和 Y 方向的跨度作为输入数据，最大层间位移比作为输出数据。最后，训练了七种机器学习 (ML) 方法来预测建筑物的损伤状况，找到了用于最高预测准确性的最佳训练建筑物、IM 和 ML 方法组合。
